| Message ID | 1671506505-104676-1-git-send-email-guwen@linux.alibaba.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | net/smc:Introduce SMC-D based loopback acceleration | expand |
On Tue, 2022-12-20 at 11:21 +0800, Wen Gu wrote: > Hi, all > > # Background > > As previously mentioned in [1], we (Alibaba Cloud) are trying to use SMC > to accelerate TCP applications in cloud environment, improving inter-host > or inter-VM communication. > > In addition of these, we also found the value of SMC-D in scenario of local > inter-process communication, such as accelerate communication between containers > within the same host. So this RFC tries to provide a SMC-D loopback solution > in such scenario, to bring a significant improvement in latency and throughput > compared to TCP loopback. > > # Design > > This patch set provides a kind of SMC-D loopback solution. > > Patch #1/5 and #2/5 provide an SMC-D based dummy device, preparing for the > inter-process communication acceleration. Except for loopback acceleration, > the dummy device can also meet the requirements mentioned in [2], which is > providing a way to test SMC-D logic for broad community without ISM device. > > +------------------------------------------+ > | +-----------+ +-----------+ | > | | process A | | process B | | > | +-----------+ +-----------+ | > | ^ ^ | > | | +---------------+ | | > | | | SMC stack | | | > | +--->| +-----------+ |<--| | > | | | dummy | | | > | | | device | | | > | +-+-----------+-+ | > | VM | > +------------------------------------------+ > > Patch #3/5, #4/5, #5/5 provides a way to avoid data copy from sndbuf to RMB > and improve SMC-D loopback performance. Through extending smcd_ops with two > new semantic: attach_dmb and detach_dmb, sender's sndbuf shares the same > physical memory region with receiver's RMB. The data copied from userspace > to sender's sndbuf directly reaches the receiver's RMB without unnecessary > memory copy in the same kernel. > > +----------+ +----------+ > | socket A | | socket B | > +----------+ +----------+ > | ^ > | +---------+ | > regard as | | ----------| > local sndbuf | B's | regard as > | | RMB | local RMB > |-------> | | > +---------+ Hi Wen Gu, I maintain the s390 specific PCI support in Linux and would like to provide a bit of background on this. You're surely wondering why we even have a copy in there for our ISM virtual PCI device. To understand why this copy operation exists and why we need to keep it working, one needs a bit of s390 aka mainframe background. On s390 all (currently supported) native machines have a mandatory machine level hypervisor. All OSs whether z/OS or Linux run either on this machine level hypervisor as so called Logical Partitions (LPARs) or as second/third/… level guests on e.g. a KVM or z/VM hypervisor that in turn runs in an LPAR. Now, in terms of memory this machine level hypervisor sometimes called PR/SM unlike KVM, z/VM, or VMWare is a partitioning hypervisor without paging. This is one of the main reasons for the very-near-native performance of the machine hypervisor as the memory of its guests acts just like native RAM on other systems. It is never paged out and always accessible to IOMMU translated DMA from devices without the need for pinning pages and besides a trivial offset/limit adjustment an LPAR's MMU does the same amount of work as an MMU on a bare metal x86_64/ARM64 box. It also means however that when SMC-D is used to communicate between LPARs via an ISM device there is no way of mapping the DMBs to the same physical memory as there exists no MMU-like layer spanning partitions that could do such a mapping. Meanwhile for machine level firmware including the ISM virtual PCI device it is still possible to _copy_ memory between different memory partitions. So yeah while I do see the appeal of skipping the memcpy() for loopback or even between guests of a paging hypervisor such as KVM, which can map the DMBs on the same physical memory, we must keep in mind this original use case requiring a copy operation. Thanks, Niklas > > # Benchmark Test > > * Test environments: > - VM with Intel Xeon Platinum 8 core 2.50GHz, 16 GiB mem. > - SMC sndbuf/RMB size 1MB. > > * Test object: > - TCP: run on TCP loopback. > - domain: run on UNIX domain. > - SMC lo: run on SMC loopback device with patch #1/5 ~ #2/5. > - SMC lo-nocpy: run on SMC loopback device with patch #1/5 ~ #5/5. > > 1. ipc-benchmark (see [3]) > > - ./<foo> -c 1000000 -s 100 > > TCP domain SMC-lo SMC-lo-nocpy > Message > rate (msg/s) 75140 129548(+72.41) 152266(+102.64%) 151914(+102.17%) Interesting that it does beat UNIX domain sockets. Also, see my below comment for nginx/wrk as this seems very similar. > > 2. sockperf > > - serv: <smc_run> taskset -c <cpu> sockperf sr --tcp > - clnt: <smc_run> taskset -c <cpu> sockperf { tp | pp } --tcp --msg-size={ 64000 for tp | 14 for pp } -i 127.0.0.1 -t 30 > > TCP SMC-lo SMC-lo-nocpy > Bandwidth(MBps) 4943.359 4936.096(-0.15%) 8239.624(+66.68%) > Latency(us) 6.372 3.359(-47.28%) 3.25(-49.00%) > > 3. iperf3 > > - serv: <smc_run> taskset -c <cpu> iperf3 -s > - clnt: <smc_run> taskset -c <cpu> iperf3 -c 127.0.0.1 -t 15 > > TCP SMC-lo SMC-lo-nocpy > Bitrate(Gb/s) 40.5 41.4(+2.22%) 76.4(+88.64%) > > 4. nginx/wrk > > - serv: <smc_run> nginx > - clnt: <smc_run> wrk -t 8 -c 500 -d 30 http://127.0.0.1:80 > > TCP SMC-lo SMC-lo-nocpy > Requests/s 154643.22 220894.03(+42.84%) 226754.3(+46.63%) This result is very interesting indeed. So with the much more realistic nginx/wrk workload it seems to copy hurts much less than the iperf3/sockperf would suggest while SMC-D itself seems to help more. I'd hope that this translates to actual applications as well. Maybe this makes SMC-D based loopback interesting even while keeping the copy, at least until we can come up with a sane way to work a no-copy variant into SMC-D? > > > # Discussion > > 1. API between SMC-D and ISM device > > As Jan mentioned in [2], IBM are working on placing an API between SMC-D > and the ISM device for easier use of different "devices" for SMC-D. > > So, considering that the introduction of attach_dmb or detach_dmb can > effectively avoid data copying from sndbuf to RMB and brings obvious > throughput advantages in inter-VM or inter-process scenarios, can the > attach/detach semantics be taken into consideration when designing the > API to make it a standard ISM device behavior? Due to the reasons explained above this behavior can't be emulated by ISM devices at least not when crossing partitions. Not sure if we can still incorporate it in the API and allow for both copying and remapping SMC-D like devices, it definitely needs careful consideration and I think also a better understanding of the benefit for real world workloads. > > Maybe our RFC of SMC-D based inter-process acceleration (this one) and > inter-VM acceleration (will coming soon, which is the update of [1]) > can provide some examples for new API design. And we are very glad to > discuss this on the mail list. > > 2. Way to select different ISM-like devices > > With the proposal of SMC-D loopback 'device' (this RFC) and incoming > device used for inter-VM acceleration as update of [1], SMC-D has more > options to choose from. So we need to consider that how to indicate > supported devices, how to determine which one to use, and their priority... Agree on this part, though it is for the SMC maintainers to decide, I think we would definitely want to be able to use any upcoming inter-VM devices on s390 possibly also in conjunction with ISM devices for communication across partitions. > > IMHO, this may require an update of CLC message and negotiation mechanism. > Again, we are very glad to discuss this with you on the mailing list. > > [1] https://lore.kernel.org/netdev/20220720170048.20806-1-tonylu@linux.alibaba.com/ > [2] https://lore.kernel.org/netdev/35d14144-28f7-6129-d6d3-ba16dae7a646@linux.ibm.com/ > [3] https://github.com/goldsborough/ipc-bench > > v1->v2 > 1. Fix some build WARNINGs complained by kernel test rebot > Reported-by: kernel test robot <lkp@intel.com> > 2. Add iperf3 test data. > > Wen Gu (5): > net/smc: introduce SMC-D loopback device > net/smc: choose loopback device in SMC-D communication > net/smc: add dmb attach and detach interface > net/smc: avoid data copy from sndbuf to peer RMB in SMC-D loopback > net/smc: logic of cursors update in SMC-D loopback connections > > include/net/smc.h | 3 + > net/smc/Makefile | 2 +- > net/smc/af_smc.c | 88 +++++++++++- > net/smc/smc_cdc.c | 59 ++++++-- > net/smc/smc_cdc.h | 1 + > net/smc/smc_clc.c | 4 +- > net/smc/smc_core.c | 62 +++++++++ > net/smc/smc_core.h | 2 + > net/smc/smc_ism.c | 39 +++++- > net/smc/smc_ism.h | 2 + > net/smc/smc_loopback.c | 358 +++++++++++++++++++++++++++++++++++++++++++++++++ > net/smc/smc_loopback.h | 63 +++++++++ > 12 files changed, 662 insertions(+), 21 deletions(-) > create mode 100644 net/smc/smc_loopback.c > create mode 100644 net/smc/smc_loopback.h >
On 2022/12/20 22:02, Niklas Schnelle wrote: > On Tue, 2022-12-20 at 11:21 +0800, Wen Gu wrote: >> Hi, all >> >> # Background >> >> As previously mentioned in [1], we (Alibaba Cloud) are trying to use SMC >> to accelerate TCP applications in cloud environment, improving inter-host >> or inter-VM communication. >> >> In addition of these, we also found the value of SMC-D in scenario of local >> inter-process communication, such as accelerate communication between containers >> within the same host. So this RFC tries to provide a SMC-D loopback solution >> in such scenario, to bring a significant improvement in latency and throughput >> compared to TCP loopback. >> >> # Design >> >> This patch set provides a kind of SMC-D loopback solution. >> >> Patch #1/5 and #2/5 provide an SMC-D based dummy device, preparing for the >> inter-process communication acceleration. Except for loopback acceleration, >> the dummy device can also meet the requirements mentioned in [2], which is >> providing a way to test SMC-D logic for broad community without ISM device. >> >> +------------------------------------------+ >> | +-----------+ +-----------+ | >> | | process A | | process B | | >> | +-----------+ +-----------+ | >> | ^ ^ | >> | | +---------------+ | | >> | | | SMC stack | | | >> | +--->| +-----------+ |<--| | >> | | | dummy | | | >> | | | device | | | >> | +-+-----------+-+ | >> | VM | >> +------------------------------------------+ >> >> Patch #3/5, #4/5, #5/5 provides a way to avoid data copy from sndbuf to RMB >> and improve SMC-D loopback performance. Through extending smcd_ops with two >> new semantic: attach_dmb and detach_dmb, sender's sndbuf shares the same >> physical memory region with receiver's RMB. The data copied from userspace >> to sender's sndbuf directly reaches the receiver's RMB without unnecessary >> memory copy in the same kernel. >> >> +----------+ +----------+ >> | socket A | | socket B | >> +----------+ +----------+ >> | ^ >> | +---------+ | >> regard as | | ----------| >> local sndbuf | B's | regard as >> | | RMB | local RMB >> |-------> | | >> +---------+ > > Hi Wen Gu, > > I maintain the s390 specific PCI support in Linux and would like to > provide a bit of background on this. You're surely wondering why we > even have a copy in there for our ISM virtual PCI device. To understand > why this copy operation exists and why we need to keep it working, one > needs a bit of s390 aka mainframe background. > > On s390 all (currently supported) native machines have a mandatory > machine level hypervisor. All OSs whether z/OS or Linux run either on > this machine level hypervisor as so called Logical Partitions (LPARs) > or as second/third/… level guests on e.g. a KVM or z/VM hypervisor that > in turn runs in an LPAR. Now, in terms of memory this machine level > hypervisor sometimes called PR/SM unlike KVM, z/VM, or VMWare is a > partitioning hypervisor without paging. This is one of the main reasons > for the very-near-native performance of the machine hypervisor as the > memory of its guests acts just like native RAM on other systems. It is > never paged out and always accessible to IOMMU translated DMA from > devices without the need for pinning pages and besides a trivial > offset/limit adjustment an LPAR's MMU does the same amount of work as > an MMU on a bare metal x86_64/ARM64 box. > > It also means however that when SMC-D is used to communicate between > LPARs via an ISM device there is no way of mapping the DMBs to the > same physical memory as there exists no MMU-like layer spanning > partitions that could do such a mapping. Meanwhile for machine level > firmware including the ISM virtual PCI device it is still possible to > _copy_ memory between different memory partitions. So yeah while I do > see the appeal of skipping the memcpy() for loopback or even between > guests of a paging hypervisor such as KVM, which can map the DMBs on > the same physical memory, we must keep in mind this original use case > requiring a copy operation. > > Thanks, > Niklas > Hi Niklas, Thank you so much for the complete and detailed explanation! This provides me a brand new perspective of s390 device that we hadn't dabbled in before. Now I understand why shared memory is unavailable between different LPARs. Our original intention of proposing loopback device and the incoming device (virtio-ism) for inter-VM is to use SMC-D to accelerate communication in the case with no existing s390 ISM devices. In our conception, s390 ISM device, loopback device and virtio-ism device are parallel and are abstracted by smcd_ops. +------------------------+ | SMC-D | +------------------------+ -------- smcd_ops --------- +------+ +------+ +------+ | s390 | | loop | |virtio| | ISM | | back | | -ism | | dev | | dev | | dev | +------+ +------+ +------+ We also believe that keeping the existing design and behavior of s390 ISM device is unshaken. What we want to get support for is some smcd_ops extension for devices with optional beneficial capability, such as nocopy here (Let's call it this for now), which is really helpful for us in inter-process and inter-VM scenario. And coincided with IBM's intention to add APIs between SMC-D and devices to support various devices for SMC-D, as mentioned in [2], we send out this RFC and the incoming virio-ism RFC, to provide some examples. >> >> # Benchmark Test >> >> * Test environments: >> - VM with Intel Xeon Platinum 8 core 2.50GHz, 16 GiB mem. >> - SMC sndbuf/RMB size 1MB. >> >> * Test object: >> - TCP: run on TCP loopback. >> - domain: run on UNIX domain. >> - SMC lo: run on SMC loopback device with patch #1/5 ~ #2/5. >> - SMC lo-nocpy: run on SMC loopback device with patch #1/5 ~ #5/5. >> >> 1. ipc-benchmark (see [3]) >> >> - ./<foo> -c 1000000 -s 100 >> >> TCP domain SMC-lo SMC-lo-nocpy >> Message >> rate (msg/s) 75140 129548(+72.41) 152266(+102.64%) 151914(+102.17%) > > Interesting that it does beat UNIX domain sockets. Also, see my below > comment for nginx/wrk as this seems very similar. > >> >> 2. sockperf >> >> - serv: <smc_run> taskset -c <cpu> sockperf sr --tcp >> - clnt: <smc_run> taskset -c <cpu> sockperf { tp | pp } --tcp --msg-size={ 64000 for tp | 14 for pp } -i 127.0.0.1 -t 30 >> >> TCP SMC-lo SMC-lo-nocpy >> Bandwidth(MBps) 4943.359 4936.096(-0.15%) 8239.624(+66.68%) >> Latency(us) 6.372 3.359(-47.28%) 3.25(-49.00%) >> >> 3. iperf3 >> >> - serv: <smc_run> taskset -c <cpu> iperf3 -s >> - clnt: <smc_run> taskset -c <cpu> iperf3 -c 127.0.0.1 -t 15 >> >> TCP SMC-lo SMC-lo-nocpy >> Bitrate(Gb/s) 40.5 41.4(+2.22%) 76.4(+88.64%) >> >> 4. nginx/wrk >> >> - serv: <smc_run> nginx >> - clnt: <smc_run> wrk -t 8 -c 500 -d 30 http://127.0.0.1:80 >> >> TCP SMC-lo SMC-lo-nocpy >> Requests/s 154643.22 220894.03(+42.84%) 226754.3(+46.63%) > > > This result is very interesting indeed. So with the much more realistic > nginx/wrk workload it seems to copy hurts much less than the > iperf3/sockperf would suggest while SMC-D itself seems to help more. > I'd hope that this translates to actual applications as well. Maybe > this makes SMC-D based loopback interesting even while keeping the > copy, at least until we can come up with a sane way to work a no-copy > variant into SMC-D? > I agree, nginx/wrk workload is much more realistic for many applications. But we also encounter many other cases similar to sockperf on the cloud, which requires high throughput, such as AI training and big data. So avoidance of copying between DMBs can help these cases a lot :) >> >> >> # Discussion >> >> 1. API between SMC-D and ISM device >> >> As Jan mentioned in [2], IBM are working on placing an API between SMC-D >> and the ISM device for easier use of different "devices" for SMC-D. >> >> So, considering that the introduction of attach_dmb or detach_dmb can >> effectively avoid data copying from sndbuf to RMB and brings obvious >> throughput advantages in inter-VM or inter-process scenarios, can the >> attach/detach semantics be taken into consideration when designing the >> API to make it a standard ISM device behavior? ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ > > Due to the reasons explained above this behavior can't be emulated by > ISM devices at least not when crossing partitions. Not sure if we can > still incorporate it in the API and allow for both copying and > remapping SMC-D like devices, it definitely needs careful consideration > and I think also a better understanding of the benefit for real world > workloads. > Here I am not rigorous. Nocopy shouldn't be a standard ISM device behavior indeed. Actually we hope it be a standard optional _SMC-D_ device behavior and defined by smcd_ops. For devices don't support these options, like ISM device on s390 architecture, .attach_dmb/.detach_dmb and other reasonable extensions (which will be proposed to discuss in incoming virtio-ism RFC) can be set to NULL or return invalid. And for devices do support, they may be used for improving performance in some cases. In addition, can I know more latest news about the API design? :) , like its scale, will it be a almost refactor of existing interface or incremental patching? and its object, will it be tailored for exact ISM behavior or to reserve some options for other devices, like nocopy here? From my understanding of [2], it might be the latter? >> >> Maybe our RFC of SMC-D based inter-process acceleration (this one) and >> inter-VM acceleration (will coming soon, which is the update of [1]) >> can provide some examples for new API design. And we are very glad to >> discuss this on the mail list. >> >> 2. Way to select different ISM-like devices >> >> With the proposal of SMC-D loopback 'device' (this RFC) and incoming >> device used for inter-VM acceleration as update of [1], SMC-D has more >> options to choose from. So we need to consider that how to indicate >> supported devices, how to determine which one to use, and their priority... > > Agree on this part, though it is for the SMC maintainers to decide, I > think we would definitely want to be able to use any upcoming inter-VM > devices on s390 possibly also in conjunction with ISM devices for > communication across partitions. > Yes, this part needs to be discussed with SMC maintainers. And thank you, we are very glad if our devices can be applied on s390 through the efforts. Best Regards, Wen Gu >> >> IMHO, this may require an update of CLC message and negotiation mechanism. >> Again, we are very glad to discuss this with you on the mailing list. >> >> [1] https://lore.kernel.org/netdev/20220720170048.20806-1-tonylu@linux.alibaba.com/ >> [2] https://lore.kernel.org/netdev/35d14144-28f7-6129-d6d3-ba16dae7a646@linux.ibm.com/ >> [3] https://github.com/goldsborough/ipc-bench >> >> v1->v2 >> 1. Fix some build WARNINGs complained by kernel test rebot >> Reported-by: kernel test robot <lkp@intel.com> >> 2. Add iperf3 test data. >> >> Wen Gu (5): >> net/smc: introduce SMC-D loopback device >> net/smc: choose loopback device in SMC-D communication >> net/smc: add dmb attach and detach interface >> net/smc: avoid data copy from sndbuf to peer RMB in SMC-D loopback >> net/smc: logic of cursors update in SMC-D loopback connections >> >> include/net/smc.h | 3 + >> net/smc/Makefile | 2 +- >> net/smc/af_smc.c | 88 +++++++++++- >> net/smc/smc_cdc.c | 59 ++++++-- >> net/smc/smc_cdc.h | 1 + >> net/smc/smc_clc.c | 4 +- >> net/smc/smc_core.c | 62 +++++++++ >> net/smc/smc_core.h | 2 + >> net/smc/smc_ism.c | 39 +++++- >> net/smc/smc_ism.h | 2 + >> net/smc/smc_loopback.c | 358 +++++++++++++++++++++++++++++++++++++++++++++++++ >> net/smc/smc_loopback.h | 63 +++++++++ >> 12 files changed, 662 insertions(+), 21 deletions(-) >> create mode 100644 net/smc/smc_loopback.c >> create mode 100644 net/smc/smc_loopback.h >>
On Tue, Dec 20, 2022 at 03:02:45PM +0100, Niklas Schnelle wrote: >On Tue, 2022-12-20 at 11:21 +0800, Wen Gu wrote: >> Hi, all >> >> # Background >> >> As previously mentioned in [1], we (Alibaba Cloud) are trying to use SMC >> to accelerate TCP applications in cloud environment, improving inter-host >> or inter-VM communication. >> >> In addition of these, we also found the value of SMC-D in scenario of local >> inter-process communication, such as accelerate communication between containers >> within the same host. So this RFC tries to provide a SMC-D loopback solution >> in such scenario, to bring a significant improvement in latency and throughput >> compared to TCP loopback. >> >> # Design >> >> This patch set provides a kind of SMC-D loopback solution. >> >> Patch #1/5 and #2/5 provide an SMC-D based dummy device, preparing for the >> inter-process communication acceleration. Except for loopback acceleration, >> the dummy device can also meet the requirements mentioned in [2], which is >> providing a way to test SMC-D logic for broad community without ISM device. >> >> +------------------------------------------+ >> | +-----------+ +-----------+ | >> | | process A | | process B | | >> | +-----------+ +-----------+ | >> | ^ ^ | >> | | +---------------+ | | >> | | | SMC stack | | | >> | +--->| +-----------+ |<--| | >> | | | dummy | | | >> | | | device | | | >> | +-+-----------+-+ | >> | VM | >> +------------------------------------------+ >> >> Patch #3/5, #4/5, #5/5 provides a way to avoid data copy from sndbuf to RMB >> and improve SMC-D loopback performance. Through extending smcd_ops with two >> new semantic: attach_dmb and detach_dmb, sender's sndbuf shares the same >> physical memory region with receiver's RMB. The data copied from userspace >> to sender's sndbuf directly reaches the receiver's RMB without unnecessary >> memory copy in the same kernel. >> >> +----------+ +----------+ >> | socket A | | socket B | >> +----------+ +----------+ >> | ^ >> | +---------+ | >> regard as | | ----------| >> local sndbuf | B's | regard as >> | | RMB | local RMB >> |-------> | | >> +---------+ > >Hi Wen Gu, > >I maintain the s390 specific PCI support in Linux and would like to >provide a bit of background on this. You're surely wondering why we >even have a copy in there for our ISM virtual PCI device. To understand >why this copy operation exists and why we need to keep it working, one >needs a bit of s390 aka mainframe background. > >On s390 all (currently supported) native machines have a mandatory >machine level hypervisor. All OSs whether z/OS or Linux run either on >this machine level hypervisor as so called Logical Partitions (LPARs) >or as second/third/… level guests on e.g. a KVM or z/VM hypervisor that >in turn runs in an LPAR. Now, in terms of memory this machine level >hypervisor sometimes called PR/SM unlike KVM, z/VM, or VMWare is a >partitioning hypervisor without paging. This is one of the main reasons >for the very-near-native performance of the machine hypervisor as the >memory of its guests acts just like native RAM on other systems. It is >never paged out and always accessible to IOMMU translated DMA from >devices without the need for pinning pages and besides a trivial >offset/limit adjustment an LPAR's MMU does the same amount of work as >an MMU on a bare metal x86_64/ARM64 box. > >It also means however that when SMC-D is used to communicate between >LPARs via an ISM device there is no way of mapping the DMBs to the >same physical memory as there exists no MMU-like layer spanning >partitions that could do such a mapping. Meanwhile for machine level >firmware including the ISM virtual PCI device it is still possible to >_copy_ memory between different memory partitions. So yeah while I do >see the appeal of skipping the memcpy() for loopback or even between >guests of a paging hypervisor such as KVM, which can map the DMBs on >the same physical memory, we must keep in mind this original use case >requiring a copy operation. > >Thanks, >Niklas > >> >> # Benchmark Test >> >> * Test environments: >> - VM with Intel Xeon Platinum 8 core 2.50GHz, 16 GiB mem. >> - SMC sndbuf/RMB size 1MB. >> >> * Test object: >> - TCP: run on TCP loopback. >> - domain: run on UNIX domain. >> - SMC lo: run on SMC loopback device with patch #1/5 ~ #2/5. >> - SMC lo-nocpy: run on SMC loopback device with patch #1/5 ~ #5/5. >> >> 1. ipc-benchmark (see [3]) >> >> - ./<foo> -c 1000000 -s 100 >> >> TCP domain SMC-lo SMC-lo-nocpy >> Message >> rate (msg/s) 75140 129548(+72.41) 152266(+102.64%) 151914(+102.17%) > >Interesting that it does beat UNIX domain sockets. Also, see my below >comment for nginx/wrk as this seems very similar. > >> >> 2. sockperf >> >> - serv: <smc_run> taskset -c <cpu> sockperf sr --tcp >> - clnt: <smc_run> taskset -c <cpu> sockperf { tp | pp } --tcp --msg-size={ 64000 for tp | 14 for pp } -i 127.0.0.1 -t 30 >> >> TCP SMC-lo SMC-lo-nocpy >> Bandwidth(MBps) 4943.359 4936.096(-0.15%) 8239.624(+66.68%) >> Latency(us) 6.372 3.359(-47.28%) 3.25(-49.00%) >> >> 3. iperf3 >> >> - serv: <smc_run> taskset -c <cpu> iperf3 -s >> - clnt: <smc_run> taskset -c <cpu> iperf3 -c 127.0.0.1 -t 15 >> >> TCP SMC-lo SMC-lo-nocpy >> Bitrate(Gb/s) 40.5 41.4(+2.22%) 76.4(+88.64%) >> >> 4. nginx/wrk >> >> - serv: <smc_run> nginx >> - clnt: <smc_run> wrk -t 8 -c 500 -d 30 http://127.0.0.1:80 >> >> TCP SMC-lo SMC-lo-nocpy >> Requests/s 154643.22 220894.03(+42.84%) 226754.3(+46.63%) > > >This result is very interesting indeed. So with the much more realistic >nginx/wrk workload it seems to copy hurts much less than the >iperf3/sockperf would suggest while SMC-D itself seems to help more. >I'd hope that this translates to actual applications as well. Maybe >this makes SMC-D based loopback interesting even while keeping the >copy, at least until we can come up with a sane way to work a no-copy >variant into SMC-D? Yes, SMC-D based loopback shows great advantages over TCP loopback, with or without copy. The advantage of zero-copy should be observed when we need to transfer a large mount of data. But here in this wrk/nginx case, the test file transferred from server to client is a small file. So we didn't see much gain. If we use a large file(e.g >=1MB file), I think we should observe a much different result. Thinks!
On 2022/12/20 11:21, Wen Gu wrote: > Hi, all > > # Background > > As previously mentioned in [1], we (Alibaba Cloud) are trying to use SMC > to accelerate TCP applications in cloud environment, improving inter-host > or inter-VM communication. > > In addition of these, we also found the value of SMC-D in scenario of local > inter-process communication, such as accelerate communication between containers > within the same host. So this RFC tries to provide a SMC-D loopback solution > in such scenario, to bring a significant improvement in latency and throughput > compared to TCP loopback. > > # Design > > This patch set provides a kind of SMC-D loopback solution. > > Patch #1/5 and #2/5 provide an SMC-D based dummy device, preparing for the > inter-process communication acceleration. Except for loopback acceleration, > the dummy device can also meet the requirements mentioned in [2], which is > providing a way to test SMC-D logic for broad community without ISM device. > > +------------------------------------------+ > | +-----------+ +-----------+ | > | | process A | | process B | | > | +-----------+ +-----------+ | > | ^ ^ | > | | +---------------+ | | > | | | SMC stack | | | > | +--->| +-----------+ |<--| | > | | | dummy | | | > | | | device | | | > | +-+-----------+-+ | > | VM | > +------------------------------------------+ > > Patch #3/5, #4/5, #5/5 provides a way to avoid data copy from sndbuf to RMB > and improve SMC-D loopback performance. Through extending smcd_ops with two > new semantic: attach_dmb and detach_dmb, sender's sndbuf shares the same > physical memory region with receiver's RMB. The data copied from userspace > to sender's sndbuf directly reaches the receiver's RMB without unnecessary > memory copy in the same kernel. > > +----------+ +----------+ > | socket A | | socket B | > +----------+ +----------+ > | ^ > | +---------+ | > regard as | | ----------| > local sndbuf | B's | regard as > | | RMB | local RMB > |-------> | | > +---------+ > > # Benchmark Test > > * Test environments: > - VM with Intel Xeon Platinum 8 core 2.50GHz, 16 GiB mem. > - SMC sndbuf/RMB size 1MB. > > * Test object: > - TCP: run on TCP loopback. > - domain: run on UNIX domain. > - SMC lo: run on SMC loopback device with patch #1/5 ~ #2/5. > - SMC lo-nocpy: run on SMC loopback device with patch #1/5 ~ #5/5. > > 1. ipc-benchmark (see [3]) > > - ./<foo> -c 1000000 -s 100 > > TCP domain SMC-lo SMC-lo-nocpy > Message > rate (msg/s) 75140 129548(+72.41) 152266(+102.64%) 151914(+102.17%) > > 2. sockperf > > - serv: <smc_run> taskset -c <cpu> sockperf sr --tcp > - clnt: <smc_run> taskset -c <cpu> sockperf { tp | pp } --tcp --msg-size={ 64000 for tp | 14 for pp } -i 127.0.0.1 -t 30 > > TCP SMC-lo SMC-lo-nocpy > Bandwidth(MBps) 4943.359 4936.096(-0.15%) 8239.624(+66.68%) > Latency(us) 6.372 3.359(-47.28%) 3.25(-49.00%) > > 3. iperf3 > > - serv: <smc_run> taskset -c <cpu> iperf3 -s > - clnt: <smc_run> taskset -c <cpu> iperf3 -c 127.0.0.1 -t 15 > > TCP SMC-lo SMC-lo-nocpy > Bitrate(Gb/s) 40.5 41.4(+2.22%) 76.4(+88.64%) > > 4. nginx/wrk > > - serv: <smc_run> nginx > - clnt: <smc_run> wrk -t 8 -c 500 -d 30 http://127.0.0.1:80 > > TCP SMC-lo SMC-lo-nocpy > Requests/s 154643.22 220894.03(+42.84%) 226754.3(+46.63%) > > > # Discussion > > 1. API between SMC-D and ISM device > > As Jan mentioned in [2], IBM are working on placing an API between SMC-D > and the ISM device for easier use of different "devices" for SMC-D. > > So, considering that the introduction of attach_dmb or detach_dmb can > effectively avoid data copying from sndbuf to RMB and brings obvious > throughput advantages in inter-VM or inter-process scenarios, can the > attach/detach semantics be taken into consideration when designing the > API to make it a standard ISM device behavior? > > Maybe our RFC of SMC-D based inter-process acceleration (this one) and > inter-VM acceleration (will coming soon, which is the update of [1]) The patch of SMC-D + virtio-ism device is now discussed in virtio community: https://lists.oasis-open.org/archives/virtio-comment/202212/msg00030.html > can provide some examples for new API design. And we are very glad to > discuss this on the mail list. > > 2. Way to select different ISM-like devices > > With the proposal of SMC-D loopback 'device' (this RFC) and incoming > device used for inter-VM acceleration as update of [1], SMC-D has more > options to choose from. So we need to consider that how to indicate > supported devices, how to determine which one to use, and their priority... > > IMHO, this may require an update of CLC message and negotiation mechanism. > Again, we are very glad to discuss this with you on the mailing list. > > [1] https://lore.kernel.org/netdev/20220720170048.20806-1-tonylu@linux.alibaba.com/ > [2] https://lore.kernel.org/netdev/35d14144-28f7-6129-d6d3-ba16dae7a646@linux.ibm.com/ > [3] https://github.com/goldsborough/ipc-bench > > v1->v2 > 1. Fix some build WARNINGs complained by kernel test rebot > Reported-by: kernel test robot <lkp@intel.com> > 2. Add iperf3 test data. > > Wen Gu (5): > net/smc: introduce SMC-D loopback device > net/smc: choose loopback device in SMC-D communication > net/smc: add dmb attach and detach interface > net/smc: avoid data copy from sndbuf to peer RMB in SMC-D loopback > net/smc: logic of cursors update in SMC-D loopback connections > > include/net/smc.h | 3 + > net/smc/Makefile | 2 +- > net/smc/af_smc.c | 88 +++++++++++- > net/smc/smc_cdc.c | 59 ++++++-- > net/smc/smc_cdc.h | 1 + > net/smc/smc_clc.c | 4 +- > net/smc/smc_core.c | 62 +++++++++ > net/smc/smc_core.h | 2 + > net/smc/smc_ism.c | 39 +++++- > net/smc/smc_ism.h | 2 + > net/smc/smc_loopback.c | 358 +++++++++++++++++++++++++++++++++++++++++++++++++ > net/smc/smc_loopback.h | 63 +++++++++ > 12 files changed, 662 insertions(+), 21 deletions(-) > create mode 100644 net/smc/smc_loopback.c > create mode 100644 net/smc/smc_loopback.h >
On 21.12.22 14:14, Wen Gu wrote: > > > On 2022/12/20 22:02, Niklas Schnelle wrote: > >> On Tue, 2022-12-20 at 11:21 +0800, Wen Gu wrote: >>> Hi, all >>> >>> # Background >>> >>> As previously mentioned in [1], we (Alibaba Cloud) are trying to use SMC >>> to accelerate TCP applications in cloud environment, improving inter-host >>> or inter-VM communication. >>> >>> In addition of these, we also found the value of SMC-D in scenario of local >>> inter-process communication, such as accelerate communication between containers >>> within the same host. So this RFC tries to provide a SMC-D loopback solution >>> in such scenario, to bring a significant improvement in latency and throughput >>> compared to TCP loopback. >>> >>> # Design >>> >>> This patch set provides a kind of SMC-D loopback solution. >>> >>> Patch #1/5 and #2/5 provide an SMC-D based dummy device, preparing for the >>> inter-process communication acceleration. Except for loopback acceleration, >>> the dummy device can also meet the requirements mentioned in [2], which is >>> providing a way to test SMC-D logic for broad community without ISM device. >>> >>> +------------------------------------------+ >>> | +-----------+ +-----------+ | >>> | | process A | | process B | | >>> | +-----------+ +-----------+ | >>> | ^ ^ | >>> | | +---------------+ | | >>> | | | SMC stack | | | >>> | +--->| +-----------+ |<--| | >>> | | | dummy | | | >>> | | | device | | | >>> | +-+-----------+-+ | >>> | VM | >>> +------------------------------------------+ >>> >>> Patch #3/5, #4/5, #5/5 provides a way to avoid data copy from sndbuf to RMB >>> and improve SMC-D loopback performance. Through extending smcd_ops with two >>> new semantic: attach_dmb and detach_dmb, sender's sndbuf shares the same >>> physical memory region with receiver's RMB. The data copied from userspace >>> to sender's sndbuf directly reaches the receiver's RMB without unnecessary >>> memory copy in the same kernel. >>> >>> +----------+ +----------+ >>> | socket A | | socket B | >>> +----------+ +----------+ >>> | ^ >>> | +---------+ | >>> regard as | | ----------| >>> local sndbuf | B's | regard as >>> | | RMB | local RMB >>> |-------> | | >>> +---------+ >> >> Hi Wen Gu, >> >> I maintain the s390 specific PCI support in Linux and would like to >> provide a bit of background on this. You're surely wondering why we >> even have a copy in there for our ISM virtual PCI device. To understand >> why this copy operation exists and why we need to keep it working, one >> needs a bit of s390 aka mainframe background. >> >> On s390 all (currently supported) native machines have a mandatory >> machine level hypervisor. All OSs whether z/OS or Linux run either on >> this machine level hypervisor as so called Logical Partitions (LPARs) >> or as second/third/… level guests on e.g. a KVM or z/VM hypervisor that >> in turn runs in an LPAR. Now, in terms of memory this machine level >> hypervisor sometimes called PR/SM unlike KVM, z/VM, or VMWare is a >> partitioning hypervisor without paging. This is one of the main reasons >> for the very-near-native performance of the machine hypervisor as the >> memory of its guests acts just like native RAM on other systems. It is >> never paged out and always accessible to IOMMU translated DMA from >> devices without the need for pinning pages and besides a trivial >> offset/limit adjustment an LPAR's MMU does the same amount of work as >> an MMU on a bare metal x86_64/ARM64 box. >> >> It also means however that when SMC-D is used to communicate between >> LPARs via an ISM device there is no way of mapping the DMBs to the >> same physical memory as there exists no MMU-like layer spanning >> partitions that could do such a mapping. Meanwhile for machine level >> firmware including the ISM virtual PCI device it is still possible to >> _copy_ memory between different memory partitions. So yeah while I do >> see the appeal of skipping the memcpy() for loopback or even between >> guests of a paging hypervisor such as KVM, which can map the DMBs on >> the same physical memory, we must keep in mind this original use case >> requiring a copy operation. >> >> Thanks, >> Niklas >> > > Hi Niklas, > > Thank you so much for the complete and detailed explanation! This provides > me a brand new perspective of s390 device that we hadn't dabbled in before. > Now I understand why shared memory is unavailable between different LPARs. > > Our original intention of proposing loopback device and the incoming device > (virtio-ism) for inter-VM is to use SMC-D to accelerate communication in the > case with no existing s390 ISM devices. In our conception, s390 ISM device, > loopback device and virtio-ism device are parallel and are abstracted by smcd_ops. > > +------------------------+ > | SMC-D | > +------------------------+ > -------- smcd_ops --------- > +------+ +------+ +------+ > | s390 | | loop | |virtio| > | ISM | | back | | -ism | > | dev | | dev | | dev | > +------+ +------+ +------+ > > We also believe that keeping the existing design and behavior of s390 ISM > device is unshaken. What we want to get support for is some smcd_ops extension > for devices with optional beneficial capability, such as nocopy here (Let's call > it this for now), which is really helpful for us in inter-process and inter-VM > scenario. > > And coincided with IBM's intention to add APIs between SMC-D and devices to > support various devices for SMC-D, as mentioned in [2], we send out this RFC and > the incoming virio-ism RFC, to provide some examples. > >>> >>> # Benchmark Test >>> >>> * Test environments: >>> - VM with Intel Xeon Platinum 8 core 2.50GHz, 16 GiB mem. >>> - SMC sndbuf/RMB size 1MB. >>> >>> * Test object: >>> - TCP: run on TCP loopback. >>> - domain: run on UNIX domain. >>> - SMC lo: run on SMC loopback device with patch #1/5 ~ #2/5. >>> - SMC lo-nocpy: run on SMC loopback device with patch #1/5 ~ #5/5. >>> >>> 1. ipc-benchmark (see [3]) >>> >>> - ./<foo> -c 1000000 -s 100 >>> >>> TCP domain SMC-lo SMC-lo-nocpy >>> Message >>> rate (msg/s) 75140 129548(+72.41) 152266(+102.64%) 151914(+102.17%) >> >> Interesting that it does beat UNIX domain sockets. Also, see my below >> comment for nginx/wrk as this seems very similar. >> >>> >>> 2. sockperf >>> >>> - serv: <smc_run> taskset -c <cpu> sockperf sr --tcp >>> - clnt: <smc_run> taskset -c <cpu> sockperf { tp | pp } --tcp --msg-size={ 64000 for tp | 14 for pp } -i 127.0.0.1 -t 30 >>> >>> TCP SMC-lo SMC-lo-nocpy >>> Bandwidth(MBps) 4943.359 4936.096(-0.15%) 8239.624(+66.68%) >>> Latency(us) 6.372 3.359(-47.28%) 3.25(-49.00%) >>> >>> 3. iperf3 >>> >>> - serv: <smc_run> taskset -c <cpu> iperf3 -s >>> - clnt: <smc_run> taskset -c <cpu> iperf3 -c 127.0.0.1 -t 15 >>> >>> TCP SMC-lo SMC-lo-nocpy >>> Bitrate(Gb/s) 40.5 41.4(+2.22%) 76.4(+88.64%) >>> >>> 4. nginx/wrk >>> >>> - serv: <smc_run> nginx >>> - clnt: <smc_run> wrk -t 8 -c 500 -d 30 http://127.0.0.1:80 >>> >>> TCP SMC-lo SMC-lo-nocpy >>> Requests/s 154643.22 220894.03(+42.84%) 226754.3(+46.63%) >> >> >> This result is very interesting indeed. So with the much more realistic >> nginx/wrk workload it seems to copy hurts much less than the >> iperf3/sockperf would suggest while SMC-D itself seems to help more. >> I'd hope that this translates to actual applications as well. Maybe >> this makes SMC-D based loopback interesting even while keeping the >> copy, at least until we can come up with a sane way to work a no-copy >> variant into SMC-D? >> > > I agree, nginx/wrk workload is much more realistic for many applications. > > But we also encounter many other cases similar to sockperf on the cloud, which > requires high throughput, such as AI training and big data. > > So avoidance of copying between DMBs can help these cases a lot :) > >>> >>> >>> # Discussion >>> >>> 1. API between SMC-D and ISM device >>> >>> As Jan mentioned in [2], IBM are working on placing an API between SMC-D >>> and the ISM device for easier use of different "devices" for SMC-D. >>> >>> So, considering that the introduction of attach_dmb or detach_dmb can >>> effectively avoid data copying from sndbuf to RMB and brings obvious >>> throughput advantages in inter-VM or inter-process scenarios, can the >>> attach/detach semantics be taken into consideration when designing the >>> API to make it a standard ISM device behavior? > ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ >> >> Due to the reasons explained above this behavior can't be emulated by >> ISM devices at least not when crossing partitions. Not sure if we can >> still incorporate it in the API and allow for both copying and >> remapping SMC-D like devices, it definitely needs careful consideration >> and I think also a better understanding of the benefit for real world >> workloads. >> > > Here I am not rigorous. > > Nocopy shouldn't be a standard ISM device behavior indeed. Actually we hope it be a > standard optional _SMC-D_ device behavior and defined by smcd_ops. > > For devices don't support these options, like ISM device on s390 architecture, > .attach_dmb/.detach_dmb and other reasonable extensions (which will be proposed to > discuss in incoming virtio-ism RFC) can be set to NULL or return invalid. And for > devices do support, they may be used for improving performance in some cases. > > In addition, can I know more latest news about the API design? :) , like its scale, will > it be a almost refactor of existing interface or incremental patching? and its object, > will it be tailored for exact ISM behavior or to reserve some options for other devices, > like nocopy here? From my understanding of [2], it might be the latter? > >>> >>> Maybe our RFC of SMC-D based inter-process acceleration (this one) and >>> inter-VM acceleration (will coming soon, which is the update of [1]) >>> can provide some examples for new API design. And we are very glad to >>> discuss this on the mail list. >>> >>> 2. Way to select different ISM-like devices >>> >>> With the proposal of SMC-D loopback 'device' (this RFC) and incoming >>> device used for inter-VM acceleration as update of [1], SMC-D has more >>> options to choose from. So we need to consider that how to indicate >>> supported devices, how to determine which one to use, and their priority... >> >> Agree on this part, though it is for the SMC maintainers to decide, I >> think we would definitely want to be able to use any upcoming inter-VM >> devices on s390 possibly also in conjunction with ISM devices for >> communication across partitions. >> > > Yes, this part needs to be discussed with SMC maintainers. And thank you, we are very glad > if our devices can be applied on s390 through the efforts. > > > Best Regards, > Wen Gu > >>> >>> IMHO, this may require an update of CLC message and negotiation mechanism. >>> Again, we are very glad to discuss this with you on the mailing list. As described in SMC protocol (including SMC-D): https://www.ibm.com/support/pages/system/files/inline-files/IBM%20Shared%20Memory%20Communications%20Version%202_2.pdf the CLC messages provide a list of up to 8 ISM devices to chose from. So I would hope that we can use the existing protocol. The challenge will be to define GID (Global Interface ID) and CHID (a fabric ID) in a meaningful way for the new devices. There is always smcd_ops->query_remote_gid() as a safety net. But the idea is that a CHID mismatch is a fast way to tell that these 2 interfaces do match. >>> >>> [1] https://lore.kernel.org/netdev/20220720170048.20806-1-tonylu@linux.alibaba.com/ >>> [2] https://lore.kernel.org/netdev/35d14144-28f7-6129-d6d3-ba16dae7a646@linux.ibm.com/ >>> [3] https://github.com/goldsborough/ipc-bench >>> >>> v1->v2 >>> 1. Fix some build WARNINGs complained by kernel test rebot >>> Reported-by: kernel test robot <lkp@intel.com> >>> 2. Add iperf3 test data. >>> >>> Wen Gu (5): >>> net/smc: introduce SMC-D loopback device >>> net/smc: choose loopback device in SMC-D communication >>> net/smc: add dmb attach and detach interface >>> net/smc: avoid data copy from sndbuf to peer RMB in SMC-D loopback >>> net/smc: logic of cursors update in SMC-D loopback connections >>> >>> include/net/smc.h | 3 + >>> net/smc/Makefile | 2 +- >>> net/smc/af_smc.c | 88 +++++++++++- >>> net/smc/smc_cdc.c | 59 ++++++-- >>> net/smc/smc_cdc.h | 1 + >>> net/smc/smc_clc.c | 4 +- >>> net/smc/smc_core.c | 62 +++++++++ >>> net/smc/smc_core.h | 2 + >>> net/smc/smc_ism.c | 39 +++++- >>> net/smc/smc_ism.h | 2 + >>> net/smc/smc_loopback.c | 358 +++++++++++++++++++++++++++++++++++++++++++++++++ >>> net/smc/smc_loopback.h | 63 +++++++++ >>> 12 files changed, 662 insertions(+), 21 deletions(-) >>> create mode 100644 net/smc/smc_loopback.c >>> create mode 100644 net/smc/smc_loopback.h >>>
On 2023/1/5 00:09, Alexandra Winter wrote: > > > On 21.12.22 14:14, Wen Gu wrote: >> >> >> On 2022/12/20 22:02, Niklas Schnelle wrote: >> >>> On Tue, 2022-12-20 at 11:21 +0800, Wen Gu wrote: >>>> Hi, all >>>> >>>> # Background >>>> >>>> As previously mentioned in [1], we (Alibaba Cloud) are trying to use SMC >>>> to accelerate TCP applications in cloud environment, improving inter-host >>>> or inter-VM communication. >>>> >>>> In addition of these, we also found the value of SMC-D in scenario of local >>>> inter-process communication, such as accelerate communication between containers >>>> within the same host. So this RFC tries to provide a SMC-D loopback solution >>>> in such scenario, to bring a significant improvement in latency and throughput >>>> compared to TCP loopback. >>>> >>>> # Design >>>> >>>> This patch set provides a kind of SMC-D loopback solution. >>>> >>>> Patch #1/5 and #2/5 provide an SMC-D based dummy device, preparing for the >>>> inter-process communication acceleration. Except for loopback acceleration, >>>> the dummy device can also meet the requirements mentioned in [2], which is >>>> providing a way to test SMC-D logic for broad community without ISM device. >>>> >>>> +------------------------------------------+ >>>> | +-----------+ +-----------+ | >>>> | | process A | | process B | | >>>> | +-----------+ +-----------+ | >>>> | ^ ^ | >>>> | | +---------------+ | | >>>> | | | SMC stack | | | >>>> | +--->| +-----------+ |<--| | >>>> | | | dummy | | | >>>> | | | device | | | >>>> | +-+-----------+-+ | >>>> | VM | >>>> +------------------------------------------+ >>>> >>>> Patch #3/5, #4/5, #5/5 provides a way to avoid data copy from sndbuf to RMB >>>> and improve SMC-D loopback performance. Through extending smcd_ops with two >>>> new semantic: attach_dmb and detach_dmb, sender's sndbuf shares the same >>>> physical memory region with receiver's RMB. The data copied from userspace >>>> to sender's sndbuf directly reaches the receiver's RMB without unnecessary >>>> memory copy in the same kernel. >>>> >>>> +----------+ +----------+ >>>> | socket A | | socket B | >>>> +----------+ +----------+ >>>> | ^ >>>> | +---------+ | >>>> regard as | | ----------| >>>> local sndbuf | B's | regard as >>>> | | RMB | local RMB >>>> |-------> | | >>>> +---------+ >>> >>> Hi Wen Gu, >>> >>> I maintain the s390 specific PCI support in Linux and would like to >>> provide a bit of background on this. You're surely wondering why we >>> even have a copy in there for our ISM virtual PCI device. To understand >>> why this copy operation exists and why we need to keep it working, one >>> needs a bit of s390 aka mainframe background. >>> >>> On s390 all (currently supported) native machines have a mandatory >>> machine level hypervisor. All OSs whether z/OS or Linux run either on >>> this machine level hypervisor as so called Logical Partitions (LPARs) >>> or as second/third/… level guests on e.g. a KVM or z/VM hypervisor that >>> in turn runs in an LPAR. Now, in terms of memory this machine level >>> hypervisor sometimes called PR/SM unlike KVM, z/VM, or VMWare is a >>> partitioning hypervisor without paging. This is one of the main reasons >>> for the very-near-native performance of the machine hypervisor as the >>> memory of its guests acts just like native RAM on other systems. It is >>> never paged out and always accessible to IOMMU translated DMA from >>> devices without the need for pinning pages and besides a trivial >>> offset/limit adjustment an LPAR's MMU does the same amount of work as >>> an MMU on a bare metal x86_64/ARM64 box. >>> >>> It also means however that when SMC-D is used to communicate between >>> LPARs via an ISM device there is no way of mapping the DMBs to the >>> same physical memory as there exists no MMU-like layer spanning >>> partitions that could do such a mapping. Meanwhile for machine level >>> firmware including the ISM virtual PCI device it is still possible to >>> _copy_ memory between different memory partitions. So yeah while I do >>> see the appeal of skipping the memcpy() for loopback or even between >>> guests of a paging hypervisor such as KVM, which can map the DMBs on >>> the same physical memory, we must keep in mind this original use case >>> requiring a copy operation. >>> >>> Thanks, >>> Niklas >>> >> >> Hi Niklas, >> >> Thank you so much for the complete and detailed explanation! This provides >> me a brand new perspective of s390 device that we hadn't dabbled in before. >> Now I understand why shared memory is unavailable between different LPARs. >> >> Our original intention of proposing loopback device and the incoming device >> (virtio-ism) for inter-VM is to use SMC-D to accelerate communication in the >> case with no existing s390 ISM devices. In our conception, s390 ISM device, >> loopback device and virtio-ism device are parallel and are abstracted by smcd_ops. >> >> +------------------------+ >> | SMC-D | >> +------------------------+ >> -------- smcd_ops --------- >> +------+ +------+ +------+ >> | s390 | | loop | |virtio| >> | ISM | | back | | -ism | >> | dev | | dev | | dev | >> +------+ +------+ +------+ >> >> We also believe that keeping the existing design and behavior of s390 ISM >> device is unshaken. What we want to get support for is some smcd_ops extension >> for devices with optional beneficial capability, such as nocopy here (Let's call >> it this for now), which is really helpful for us in inter-process and inter-VM >> scenario. >> >> And coincided with IBM's intention to add APIs between SMC-D and devices to >> support various devices for SMC-D, as mentioned in [2], we send out this RFC and >> the incoming virio-ism RFC, to provide some examples. >> >>>> >>>> # Benchmark Test >>>> >>>> * Test environments: >>>> - VM with Intel Xeon Platinum 8 core 2.50GHz, 16 GiB mem. >>>> - SMC sndbuf/RMB size 1MB. >>>> >>>> * Test object: >>>> - TCP: run on TCP loopback. >>>> - domain: run on UNIX domain. >>>> - SMC lo: run on SMC loopback device with patch #1/5 ~ #2/5. >>>> - SMC lo-nocpy: run on SMC loopback device with patch #1/5 ~ #5/5. >>>> >>>> 1. ipc-benchmark (see [3]) >>>> >>>> - ./<foo> -c 1000000 -s 100 >>>> >>>> TCP domain SMC-lo SMC-lo-nocpy >>>> Message >>>> rate (msg/s) 75140 129548(+72.41) 152266(+102.64%) 151914(+102.17%) >>> >>> Interesting that it does beat UNIX domain sockets. Also, see my below >>> comment for nginx/wrk as this seems very similar. >>> >>>> >>>> 2. sockperf >>>> >>>> - serv: <smc_run> taskset -c <cpu> sockperf sr --tcp >>>> - clnt: <smc_run> taskset -c <cpu> sockperf { tp | pp } --tcp --msg-size={ 64000 for tp | 14 for pp } -i 127.0.0.1 -t 30 >>>> >>>> TCP SMC-lo SMC-lo-nocpy >>>> Bandwidth(MBps) 4943.359 4936.096(-0.15%) 8239.624(+66.68%) >>>> Latency(us) 6.372 3.359(-47.28%) 3.25(-49.00%) >>>> >>>> 3. iperf3 >>>> >>>> - serv: <smc_run> taskset -c <cpu> iperf3 -s >>>> - clnt: <smc_run> taskset -c <cpu> iperf3 -c 127.0.0.1 -t 15 >>>> >>>> TCP SMC-lo SMC-lo-nocpy >>>> Bitrate(Gb/s) 40.5 41.4(+2.22%) 76.4(+88.64%) >>>> >>>> 4. nginx/wrk >>>> >>>> - serv: <smc_run> nginx >>>> - clnt: <smc_run> wrk -t 8 -c 500 -d 30 http://127.0.0.1:80 >>>> >>>> TCP SMC-lo SMC-lo-nocpy >>>> Requests/s 154643.22 220894.03(+42.84%) 226754.3(+46.63%) >>> >>> >>> This result is very interesting indeed. So with the much more realistic >>> nginx/wrk workload it seems to copy hurts much less than the >>> iperf3/sockperf would suggest while SMC-D itself seems to help more. >>> I'd hope that this translates to actual applications as well. Maybe >>> this makes SMC-D based loopback interesting even while keeping the >>> copy, at least until we can come up with a sane way to work a no-copy >>> variant into SMC-D? >>> >> >> I agree, nginx/wrk workload is much more realistic for many applications. >> >> But we also encounter many other cases similar to sockperf on the cloud, which >> requires high throughput, such as AI training and big data. >> >> So avoidance of copying between DMBs can help these cases a lot :) >> >>>> >>>> >>>> # Discussion >>>> >>>> 1. API between SMC-D and ISM device >>>> >>>> As Jan mentioned in [2], IBM are working on placing an API between SMC-D >>>> and the ISM device for easier use of different "devices" for SMC-D. >>>> >>>> So, considering that the introduction of attach_dmb or detach_dmb can >>>> effectively avoid data copying from sndbuf to RMB and brings obvious >>>> throughput advantages in inter-VM or inter-process scenarios, can the >>>> attach/detach semantics be taken into consideration when designing the >>>> API to make it a standard ISM device behavior? >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ >>> >>> Due to the reasons explained above this behavior can't be emulated by >>> ISM devices at least not when crossing partitions. Not sure if we can >>> still incorporate it in the API and allow for both copying and >>> remapping SMC-D like devices, it definitely needs careful consideration >>> and I think also a better understanding of the benefit for real world >>> workloads. >>> >> >> Here I am not rigorous. >> >> Nocopy shouldn't be a standard ISM device behavior indeed. Actually we hope it be a >> standard optional _SMC-D_ device behavior and defined by smcd_ops. >> >> For devices don't support these options, like ISM device on s390 architecture, >> .attach_dmb/.detach_dmb and other reasonable extensions (which will be proposed to >> discuss in incoming virtio-ism RFC) can be set to NULL or return invalid. And for >> devices do support, they may be used for improving performance in some cases. >> >> In addition, can I know more latest news about the API design? :) , like its scale, will >> it be a almost refactor of existing interface or incremental patching? and its object, >> will it be tailored for exact ISM behavior or to reserve some options for other devices, >> like nocopy here? From my understanding of [2], it might be the latter? >> >>>> >>>> Maybe our RFC of SMC-D based inter-process acceleration (this one) and >>>> inter-VM acceleration (will coming soon, which is the update of [1]) >>>> can provide some examples for new API design. And we are very glad to >>>> discuss this on the mail list. >>>> >>>> 2. Way to select different ISM-like devices >>>> >>>> With the proposal of SMC-D loopback 'device' (this RFC) and incoming >>>> device used for inter-VM acceleration as update of [1], SMC-D has more >>>> options to choose from. So we need to consider that how to indicate >>>> supported devices, how to determine which one to use, and their priority... >>> >>> Agree on this part, though it is for the SMC maintainers to decide, I >>> think we would definitely want to be able to use any upcoming inter-VM >>> devices on s390 possibly also in conjunction with ISM devices for >>> communication across partitions. >>> >> >> Yes, this part needs to be discussed with SMC maintainers. And thank you, we are very glad >> if our devices can be applied on s390 through the efforts. >> >> >> Best Regards, >> Wen Gu >> >>>> >>>> IMHO, this may require an update of CLC message and negotiation mechanism. >>>> Again, we are very glad to discuss this with you on the mailing list. > > As described in > SMC protocol (including SMC-D): https://www.ibm.com/support/pages/system/files/inline-files/IBM%20Shared%20Memory%20Communications%20Version%202_2.pdf > the CLC messages provide a list of up to 8 ISM devices to chose from. > So I would hope that we can use the existing protocol. > > The challenge will be to define GID (Global Interface ID) and CHID (a fabric ID) in > a meaningful way for the new devices. > There is always smcd_ops->query_remote_gid() as a safety net. But the idea is that > a CHID mismatch is a fast way to tell that these 2 interfaces do match. > > Hi Winter and all, Thanks for your reply and suggestions! And sorry for my late reply because it took me some time to understand SMC-Dv2 protocol and implementation. I agree with your opinion. The existing SMC-Dv2 protocol whose CLC messages include ism_dev[] list can solve the devices negotiation problem. And I am very willing to use the existing protocol, because we all know that the protocol update is a long and complex process. If I understand correctly, SMC-D loopback(dummy) device can coordinate with existing SMC-Dv2 protocol as follows. If there is any mistake, please point out. # Initialization - Initialize the loopback device with unique GID [Q-1]. - Register the loopback device as SMC-Dv2-capable device with a system_eid whose 24th or 28th byte is non-zero [Q-2], so that this system's smc_ism_v2_capable will be set to TRUE and SMC-Dv2 is available. # Proposal - Find the loopback device from the smcd_dev_list in smc_find_ism_v2_device_clnt(); - Record the SEID, GID and CHID[Q-3] of loopback device in the v2 extension part of CLC proposal message. # Accept - Check the GID/CHID list and SEID in CLC proposal message, and find local matched ISM device from smcd_dev_list in smc_find_ism_v2_device_serv(). If both sides of the communication are in the same VM and share the same loopback device, the SEID, GID and CHID will match and loopback device will be chosen [Q-4]. - Record the loopback device's GID/CHID and matched SEID into CLC accept message. # Confirm - Confirm the server-selected device (loopback device) accordingto CLC accept messages. - Record the loopback device's GID/CHID and server-selected SEID in CLC confirm message. Follow the above process, I supplement a patch based on this RFC in the email attachment. With the attachment patch, SMC-D loopback will switch to use SMC-Dv2 protocol. And in the above process, there are something I want to consult and discuss, which is marked with '[Q-*]' in the above description. # [Q-1]: The GID of loopback device is randomly generated in this RFC patch set, but I will find a way to unique the GID in formal patches. Any suggestions are welcome. # [Q-2]: In Linux implementation, the system_eid of the first registered smcd device will determinate system's smc_ism_v2_capable (see smcd_register_dev()). And I wonder that 1) How to define the system_eid? It can be inferred from the code that the 24th and 28th byte are special for SMC-Dv2. So in attachment patch, I define the loopback device SEID as static struct smc_lo_systemeid LO_SYSTEM_EID = { .seid_string = "SMC-SYSZ-LOSEID000000000", .serial_number = "1000", .type = "1000", }; Is there anything else I need to pay attention to? 2) Seems only the first added smcd device determinate the system smc_ism_v2_capable? If two different smcd devices respectively with v1-indicated and v2-indicated system_eid, will the order in which they are registered affects the result of smc_ism_v2_capable ? # [Q-3]: In attachment patch, I define a special CHID (0xFFFF) for loopback device, as a kind of 'unassociated ISM CHID' that not associated with any IP (OSA or HiperSockets) interfaces. What's your opinion about this? # [Q-4]: In current Linux implementation, server will select the first successfully initialized device from the candidates as the final selected one in smc_find_ism_v2_device_serv(). for (i = 0; i < matches; i++) { ini->smcd_version = SMC_V2; ini->is_smcd = true; ini->ism_selected = i; rc = smc_listen_ism_init(new_smc, ini); if (rc) { smc_find_ism_store_rc(rc, ini); /* try next active ISM device */ continue; } return; /* matching and usable V2 ISM device found */ } IMHO, maybe candidate devices should have different priorities? For example, the loopback device may be preferred to use if loopback is available. Best Regards, Wen Gu >>>> >>>> [1] https://lore.kernel.org/netdev/20220720170048.20806-1-tonylu@linux.alibaba.com/ >>>> [2] https://lore.kernel.org/netdev/35d14144-28f7-6129-d6d3-ba16dae7a646@linux.ibm.com/ >>>> [3] https://github.com/goldsborough/ipc-bench >>>> >>>> v1->v2 >>>> 1. Fix some build WARNINGs complained by kernel test rebot >>>> Reported-by: kernel test robot <lkp@intel.com> >>>> 2. Add iperf3 test data. >>>> >>>> Wen Gu (5): >>>> net/smc: introduce SMC-D loopback device >>>> net/smc: choose loopback device in SMC-D communication >>>> net/smc: add dmb attach and detach interface >>>> net/smc: avoid data copy from sndbuf to peer RMB in SMC-D loopback >>>> net/smc: logic of cursors update in SMC-D loopback connections >>>> >>>> include/net/smc.h | 3 + >>>> net/smc/Makefile | 2 +- >>>> net/smc/af_smc.c | 88 +++++++++++- >>>> net/smc/smc_cdc.c | 59 ++++++-- >>>> net/smc/smc_cdc.h | 1 + >>>> net/smc/smc_clc.c | 4 +- >>>> net/smc/smc_core.c | 62 +++++++++ >>>> net/smc/smc_core.h | 2 + >>>> net/smc/smc_ism.c | 39 +++++- >>>> net/smc/smc_ism.h | 2 + >>>> net/smc/smc_loopback.c | 358 +++++++++++++++++++++++++++++++++++++++++++++++++ >>>> net/smc/smc_loopback.h | 63 +++++++++ >>>> 12 files changed, 662 insertions(+), 21 deletions(-) >>>> create mode 100644 net/smc/smc_loopback.c >>>> create mode 100644 net/smc/smc_loopback.h >>>> From bc94984d599e2e8cbc408c42896973745c533bb7 Mon Sep 17 00:00:00 2001 From: Wen Gu <guwen@linux.alibaba.com> Date: Sat, 7 Jan 2023 16:58:37 +0800 Subject: [PATCH] net/smc: define SEID and CHID of loopback device This patch defines SEID and CHID of loopback device and take it as SMC-Dv2 device. Besides, this patch delete the most logic of RFC patch 2/5 as well because device selection will be covered by SMC-Dv2 protocol. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> --- net/smc/af_smc.c | 50 +++++--------------------------------------------- net/smc/smc_clc.c | 4 +--- net/smc/smc_loopback.c | 11 +++++++---- 3 files changed, 13 insertions(+), 52 deletions(-) diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c index c7de566..4396392 100644 --- a/net/smc/af_smc.c +++ b/net/smc/af_smc.c @@ -979,28 +979,6 @@ static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini) return 0; } -/* check if there is a lo device available for this connection. */ -static int smc_find_lo_device(struct smc_sock *smc, struct smc_init_info *ini) -{ - struct smcd_dev *sdev; - - mutex_lock(&smcd_dev_list.mutex); - list_for_each_entry(sdev, &smcd_dev_list.list, list) { - if (sdev->is_loopback && !sdev->going_away && - (!ini->ism_peer_gid[0] || - !smc_ism_cantalk(ini->ism_peer_gid[0], ini->vlan_id, - sdev))) { - ini->ism_dev[0] = sdev; - break; - } - } - mutex_unlock(&smcd_dev_list.mutex); - if (!ini->ism_dev[0]) - return SMC_CLC_DECL_NOSMCDDEV; - ini->ism_chid[0] = smc_ism_get_chid(ini->ism_dev[0]); - return 0; -} - /* is chid unique for the ism devices that are already determined? */ static bool smc_find_ism_v2_is_unique_chid(u16 chid, struct smc_init_info *ini, int cnt) @@ -1066,19 +1044,10 @@ static int smc_find_proposal_devices(struct smc_sock *smc, { int rc = 0; - /* TODO: - * How to indicate to peer if ism device and loopback - * device are both available ? - * - * The RFC patch hasn't resolved this, just simply always - * chooses loopback device first, and fallback if loopback - * communication is impossible. - */ /* check if there is an ism or loopback device available */ if (!(ini->smcd_version & SMC_V1) || - (smc_find_lo_device(smc, ini) && - (smc_find_ism_device(smc, ini) || - smc_connect_ism_vlan_setup(smc, ini)))) + smc_find_ism_device(smc, ini) || + smc_connect_ism_vlan_setup(smc, ini)) ini->smcd_version &= ~SMC_V1; /* else ISM V1 is supported for this connection */ @@ -2178,18 +2147,9 @@ static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc, ini->is_smcd = true; /* prepare ISM check */ ini->ism_peer_gid[0] = ntohll(pclc_smcd->ism.gid); - /* TODO: - * How to know that peer has both loopback and ism device ? - * - * The RFC patch hasn't resolved this, simply tries loopback - * device first, then ism device. - */ - /* find available loopback or ism device */ - if (smc_find_lo_device(new_smc, ini)) { - rc = smc_find_ism_device(new_smc, ini); - if (rc) - goto not_found; - } + rc = smc_find_ism_device(new_smc, ini); + if (rc) + goto not_found; ini->ism_selected = 0; rc = smc_listen_ism_init(new_smc, ini); diff --git a/net/smc/smc_clc.c b/net/smc/smc_clc.c index 3887692..dfb9797 100644 --- a/net/smc/smc_clc.c +++ b/net/smc/smc_clc.c @@ -486,9 +486,7 @@ static int smc_clc_prfx_set4_rcu(struct dst_entry *dst, __be32 ipv4, return -ENODEV; in_dev_for_each_ifa_rcu(ifa, in_dev) { - /* add loopback support */ - if (inet_addr_type(dev_net(dst->dev), ipv4) != RTN_LOCAL && - !inet_ifa_match(ipv4, ifa)) + if (!inet_ifa_match(ipv4, ifa)) continue; prop->prefix_len = inet_mask_len(ifa->ifa_mask); prop->outgoing_subnet = ifa->ifa_address & ifa->ifa_mask; diff --git a/net/smc/smc_loopback.c b/net/smc/smc_loopback.c index 3dedcc4..642b241 100644 --- a/net/smc/smc_loopback.c +++ b/net/smc/smc_loopback.c @@ -19,13 +19,14 @@ #include "smc_loopback.h" #define DRV_NAME "smc_lodev" +#define LO_CHID 0xFFFF /* specific for lo dev */ struct smc_lo_dev *lo_dev; static struct smc_lo_systemeid LO_SYSTEM_EID = { .seid_string = "SMC-SYSZ-LOSEID000000000", - .serial_number = "0000", - .type = "0000", + .serial_number = "1000", + .type = "1000", }; static int smc_lo_query_rgid(struct smcd_dev *smcd, u64 rgid, u32 vid_valid, @@ -33,7 +34,9 @@ static int smc_lo_query_rgid(struct smcd_dev *smcd, u64 rgid, u32 vid_valid, { struct smc_lo_dev *ldev = smcd->priv; - /* return local gid */ + if (!vid_valid || vid != ISM_RESERVED_VLANID) + return -EINVAL; + /* rgid should be equal to lgid */ if (!ldev || rgid != ldev->lgid) return -ENETUNREACH; return 0; @@ -255,7 +258,7 @@ static u8 *smc_lo_get_system_eid(void) static u16 smc_lo_get_chid(struct smcd_dev *smcd) { - return 0; + return LO_CHID; } static const struct smcd_ops lo_ops = {
On 12.01.23 13:12, Wen Gu wrote: > > > On 2023/1/5 00:09, Alexandra Winter wrote: >> >> >> On 21.12.22 14:14, Wen Gu wrote: >>> >>> >>> On 2022/12/20 22:02, Niklas Schnelle wrote: >>> >>>> On Tue, 2022-12-20 at 11:21 +0800, Wen Gu wrote: >>>>> Hi, all >>>>> >>>>> # Background >>>>> >>>>> As previously mentioned in [1], we (Alibaba Cloud) are trying to >>>>> use SMC >>>>> to accelerate TCP applications in cloud environment, improving >>>>> inter-host >>>>> or inter-VM communication. >>>>> >>>>> In addition of these, we also found the value of SMC-D in scenario >>>>> of local >>>>> inter-process communication, such as accelerate communication >>>>> between containers >>>>> within the same host. So this RFC tries to provide a SMC-D loopback >>>>> solution >>>>> in such scenario, to bring a significant improvement in latency and >>>>> throughput >>>>> compared to TCP loopback. >>>>> >>>>> # Design >>>>> >>>>> This patch set provides a kind of SMC-D loopback solution. >>>>> >>>>> Patch #1/5 and #2/5 provide an SMC-D based dummy device, preparing >>>>> for the >>>>> inter-process communication acceleration. Except for loopback >>>>> acceleration, >>>>> the dummy device can also meet the requirements mentioned in [2], >>>>> which is >>>>> providing a way to test SMC-D logic for broad community without ISM >>>>> device. >>>>> >>>>> +------------------------------------------+ >>>>> | +-----------+ +-----------+ | >>>>> | | process A | | process B | | >>>>> | +-----------+ +-----------+ | >>>>> | ^ ^ | >>>>> | | +---------------+ | | >>>>> | | | SMC stack | | | >>>>> | +--->| +-----------+ |<--| | >>>>> | | | dummy | | | >>>>> | | | device | | | >>>>> | +-+-----------+-+ | >>>>> | VM | >>>>> +------------------------------------------+ >>>>> >>>>> Patch #3/5, #4/5, #5/5 provides a way to avoid data copy from >>>>> sndbuf to RMB >>>>> and improve SMC-D loopback performance. Through extending smcd_ops >>>>> with two >>>>> new semantic: attach_dmb and detach_dmb, sender's sndbuf shares the >>>>> same >>>>> physical memory region with receiver's RMB. The data copied from >>>>> userspace >>>>> to sender's sndbuf directly reaches the receiver's RMB without >>>>> unnecessary >>>>> memory copy in the same kernel. >>>>> >>>>> +----------+ +----------+ >>>>> | socket A | | socket B | >>>>> +----------+ +----------+ >>>>> | ^ >>>>> | +---------+ | >>>>> regard as | | ----------| >>>>> local sndbuf | B's | regard as >>>>> | | RMB | local RMB >>>>> |-------> | | >>>>> +---------+ >>>> >>>> Hi Wen Gu, >>>> >>>> I maintain the s390 specific PCI support in Linux and would like to >>>> provide a bit of background on this. You're surely wondering why we >>>> even have a copy in there for our ISM virtual PCI device. To understand >>>> why this copy operation exists and why we need to keep it working, one >>>> needs a bit of s390 aka mainframe background. >>>> >>>> On s390 all (currently supported) native machines have a mandatory >>>> machine level hypervisor. All OSs whether z/OS or Linux run either on >>>> this machine level hypervisor as so called Logical Partitions (LPARs) >>>> or as second/third/… level guests on e.g. a KVM or z/VM hypervisor that >>>> in turn runs in an LPAR. Now, in terms of memory this machine level >>>> hypervisor sometimes called PR/SM unlike KVM, z/VM, or VMWare is a >>>> partitioning hypervisor without paging. This is one of the main reasons >>>> for the very-near-native performance of the machine hypervisor as the >>>> memory of its guests acts just like native RAM on other systems. It is >>>> never paged out and always accessible to IOMMU translated DMA from >>>> devices without the need for pinning pages and besides a trivial >>>> offset/limit adjustment an LPAR's MMU does the same amount of work as >>>> an MMU on a bare metal x86_64/ARM64 box. >>>> >>>> It also means however that when SMC-D is used to communicate between >>>> LPARs via an ISM device there is no way of mapping the DMBs to the >>>> same physical memory as there exists no MMU-like layer spanning >>>> partitions that could do such a mapping. Meanwhile for machine level >>>> firmware including the ISM virtual PCI device it is still possible to >>>> _copy_ memory between different memory partitions. So yeah while I do >>>> see the appeal of skipping the memcpy() for loopback or even between >>>> guests of a paging hypervisor such as KVM, which can map the DMBs on >>>> the same physical memory, we must keep in mind this original use case >>>> requiring a copy operation. >>>> >>>> Thanks, >>>> Niklas >>>> >>> >>> Hi Niklas, >>> >>> Thank you so much for the complete and detailed explanation! This >>> provides >>> me a brand new perspective of s390 device that we hadn't dabbled in >>> before. >>> Now I understand why shared memory is unavailable between different >>> LPARs. >>> >>> Our original intention of proposing loopback device and the incoming >>> device >>> (virtio-ism) for inter-VM is to use SMC-D to accelerate communication >>> in the >>> case with no existing s390 ISM devices. In our conception, s390 ISM >>> device, >>> loopback device and virtio-ism device are parallel and are abstracted >>> by smcd_ops. >>> >>> +------------------------+ >>> | SMC-D | >>> +------------------------+ >>> -------- smcd_ops --------- >>> +------+ +------+ +------+ >>> | s390 | | loop | |virtio| >>> | ISM | | back | | -ism | >>> | dev | | dev | | dev | >>> +------+ +------+ +------+ >>> >>> We also believe that keeping the existing design and behavior of s390 >>> ISM >>> device is unshaken. What we want to get support for is some smcd_ops >>> extension >>> for devices with optional beneficial capability, such as nocopy here >>> (Let's call >>> it this for now), which is really helpful for us in inter-process and >>> inter-VM >>> scenario. >>> >>> And coincided with IBM's intention to add APIs between SMC-D and >>> devices to >>> support various devices for SMC-D, as mentioned in [2], we send out >>> this RFC and >>> the incoming virio-ism RFC, to provide some examples. >>> >>>>> >>>>> # Benchmark Test >>>>> >>>>> * Test environments: >>>>> - VM with Intel Xeon Platinum 8 core 2.50GHz, 16 GiB mem. >>>>> - SMC sndbuf/RMB size 1MB. >>>>> >>>>> * Test object: >>>>> - TCP: run on TCP loopback. >>>>> - domain: run on UNIX domain. >>>>> - SMC lo: run on SMC loopback device with patch #1/5 ~ #2/5. >>>>> - SMC lo-nocpy: run on SMC loopback device with patch #1/5 >>>>> ~ #5/5. >>>>> >>>>> 1. ipc-benchmark (see [3]) >>>>> >>>>> - ./<foo> -c 1000000 -s 100 >>>>> >>>>> TCP domain >>>>> SMC-lo SMC-lo-nocpy >>>>> Message >>>>> rate (msg/s) 75140 129548(+72.41) >>>>> 152266(+102.64%) 151914(+102.17%) >>>> >>>> Interesting that it does beat UNIX domain sockets. Also, see my below >>>> comment for nginx/wrk as this seems very similar. >>>> >>>>> >>>>> 2. sockperf >>>>> >>>>> - serv: <smc_run> taskset -c <cpu> sockperf sr --tcp >>>>> - clnt: <smc_run> taskset -c <cpu> sockperf { tp | pp } --tcp >>>>> --msg-size={ 64000 for tp | 14 for pp } -i 127.0.0.1 -t 30 >>>>> >>>>> TCP SMC-lo >>>>> SMC-lo-nocpy >>>>> Bandwidth(MBps) 4943.359 4936.096(-0.15%) >>>>> 8239.624(+66.68%) >>>>> Latency(us) 6.372 3.359(-47.28%) >>>>> 3.25(-49.00%) >>>>> >>>>> 3. iperf3 >>>>> >>>>> - serv: <smc_run> taskset -c <cpu> iperf3 -s >>>>> - clnt: <smc_run> taskset -c <cpu> iperf3 -c 127.0.0.1 -t 15 >>>>> >>>>> TCP SMC-lo >>>>> SMC-lo-nocpy >>>>> Bitrate(Gb/s) 40.5 41.4(+2.22%) >>>>> 76.4(+88.64%) >>>>> >>>>> 4. nginx/wrk >>>>> >>>>> - serv: <smc_run> nginx >>>>> - clnt: <smc_run> wrk -t 8 -c 500 -d 30 http://127.0.0.1:80 >>>>> >>>>> TCP SMC-lo >>>>> SMC-lo-nocpy >>>>> Requests/s 154643.22 220894.03(+42.84%) >>>>> 226754.3(+46.63%) >>>> >>>> >>>> This result is very interesting indeed. So with the much more realistic >>>> nginx/wrk workload it seems to copy hurts much less than the >>>> iperf3/sockperf would suggest while SMC-D itself seems to help more. >>>> I'd hope that this translates to actual applications as well. Maybe >>>> this makes SMC-D based loopback interesting even while keeping the >>>> copy, at least until we can come up with a sane way to work a no-copy >>>> variant into SMC-D? >>>> >>> >>> I agree, nginx/wrk workload is much more realistic for many >>> applications. >>> >>> But we also encounter many other cases similar to sockperf on the >>> cloud, which >>> requires high throughput, such as AI training and big data. >>> >>> So avoidance of copying between DMBs can help these cases a lot :) >>> >>>>> >>>>> >>>>> # Discussion >>>>> >>>>> 1. API between SMC-D and ISM device >>>>> >>>>> As Jan mentioned in [2], IBM are working on placing an API between >>>>> SMC-D >>>>> and the ISM device for easier use of different "devices" for SMC-D. >>>>> >>>>> So, considering that the introduction of attach_dmb or detach_dmb can >>>>> effectively avoid data copying from sndbuf to RMB and brings obvious >>>>> throughput advantages in inter-VM or inter-process scenarios, can the >>>>> attach/detach semantics be taken into consideration when designing the >>>>> API to make it a standard ISM device behavior? >>> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ >>>> >>>> Due to the reasons explained above this behavior can't be emulated by >>>> ISM devices at least not when crossing partitions. Not sure if we can >>>> still incorporate it in the API and allow for both copying and >>>> remapping SMC-D like devices, it definitely needs careful consideration >>>> and I think also a better understanding of the benefit for real world >>>> workloads. >>>> >>> >>> Here I am not rigorous. >>> >>> Nocopy shouldn't be a standard ISM device behavior indeed. Actually >>> we hope it be a >>> standard optional _SMC-D_ device behavior and defined by smcd_ops. >>> >>> For devices don't support these options, like ISM device on s390 >>> architecture, >>> .attach_dmb/.detach_dmb and other reasonable extensions (which will >>> be proposed to >>> discuss in incoming virtio-ism RFC) can be set to NULL or return >>> invalid. And for >>> devices do support, they may be used for improving performance in >>> some cases. >>> >>> In addition, can I know more latest news about the API design? :) , >>> like its scale, will >>> it be a almost refactor of existing interface or incremental >>> patching? and its object, >>> will it be tailored for exact ISM behavior or to reserve some options >>> for other devices, >>> like nocopy here? From my understanding of [2], it might be the latter? >>> >>>>> >>>>> Maybe our RFC of SMC-D based inter-process acceleration (this one) and >>>>> inter-VM acceleration (will coming soon, which is the update of [1]) >>>>> can provide some examples for new API design. And we are very glad to >>>>> discuss this on the mail list. >>>>> >>>>> 2. Way to select different ISM-like devices >>>>> >>>>> With the proposal of SMC-D loopback 'device' (this RFC) and incoming >>>>> device used for inter-VM acceleration as update of [1], SMC-D has more >>>>> options to choose from. So we need to consider that how to indicate >>>>> supported devices, how to determine which one to use, and their >>>>> priority... >>>> >>>> Agree on this part, though it is for the SMC maintainers to decide, I >>>> think we would definitely want to be able to use any upcoming inter-VM >>>> devices on s390 possibly also in conjunction with ISM devices for >>>> communication across partitions. >>>> >>> >>> Yes, this part needs to be discussed with SMC maintainers. And thank >>> you, we are very glad >>> if our devices can be applied on s390 through the efforts. >>> >>> >>> Best Regards, >>> Wen Gu >>> >>>>> >>>>> IMHO, this may require an update of CLC message and negotiation >>>>> mechanism. >>>>> Again, we are very glad to discuss this with you on the mailing list. >> >> As described in >> SMC protocol (including SMC-D): >> https://www.ibm.com/support/pages/system/files/inline-files/IBM%20Shared%20Memory%20Communications%20Version%202_2.pdf >> the CLC messages provide a list of up to 8 ISM devices to chose from. >> So I would hope that we can use the existing protocol. >> >> The challenge will be to define GID (Global Interface ID) and CHID (a >> fabric ID) in >> a meaningful way for the new devices. >> There is always smcd_ops->query_remote_gid() as a safety net. But the >> idea is that >> a CHID mismatch is a fast way to tell that these 2 interfaces do match. >> >> > FYI, we just sent the rest part of the API to the net-next https://lore.kernel.org/netdev/20230116092712.10176-1-jaka@linux.ibm.com/T/#t, which should answer some questions in your patch series. > Hi Winter and all, > > Thanks for your reply and suggestions! And sorry for my late reply > because it took me > some time to understand SMC-Dv2 protocol and implementation. > > I agree with your opinion. The existing SMC-Dv2 protocol whose CLC > messages include > ism_dev[] list can solve the devices negotiation problem. And I am very > willing to use > the existing protocol, because we all know that the protocol update is a > long and complex > process. > > If I understand correctly, SMC-D loopback(dummy) device can coordinate > with existing > SMC-Dv2 protocol as follows. If there is any mistake, please point out. > > > # Initialization > > - Initialize the loopback device with unique GID [Q-1]. > > - Register the loopback device as SMC-Dv2-capable device with a > system_eid whose 24th > or 28th byte is non-zero [Q-2], so that this system's > smc_ism_v2_capable will be set > to TRUE and SMC-Dv2 is available. > The decision point is the VLAN_ID, if it is x1FFF, the device will support V2. i.e. If you can have subnet with VLAN_ID x1FFF, then the SEID is necessary, so that the series or types is non-zero. (*1) > > # Proposal > > - Find the loopback device from the smcd_dev_list in > smc_find_ism_v2_device_clnt(); > > - Record the SEID, GID and CHID[Q-3] of loopback device in the v2 > extension part of CLC > proposal message. > > > # Accept > > - Check the GID/CHID list and SEID in CLC proposal message, and find > local matched ISM > device from smcd_dev_list in smc_find_ism_v2_device_serv(). If both > sides of the > communication are in the same VM and share the same loopback device, > the SEID, GID and > CHID will match and loopback device will be chosen [Q-4]. > > - Record the loopback device's GID/CHID and matched SEID into CLC accept > message. > > > # Confirm > > - Confirm the server-selected device (loopback device) accordingto CLC > accept messages. > > - Record the loopback device's GID/CHID and server-selected SEID in CLC > confirm message. > > > Follow the above process, I supplement a patch based on this RFC in the > email attachment. > With the attachment patch, SMC-D loopback will switch to use SMC-Dv2 > protocol. > > > > And in the above process, there are something I want to consult and > discuss, which is marked > with '[Q-*]' in the above description. > > # [Q-1]: > > The GID of loopback device is randomly generated in this RFC patch set, > but I will find a way > to unique the GID in formal patches. Any suggestions are welcome. > I think the randowmly generated GID is fine in your case, which is equivalent to the IP address. > > # [Q-2]: > > In Linux implementation, the system_eid of the first registered smcd > device will determinate > system's smc_ism_v2_capable (see smcd_register_dev()). > > And I wonder that > > 1) How to define the system_eid? It can be inferred from the code that > the 24th and 28th byte > are special for SMC-Dv2. So in attachment patch, I define the > loopback device SEID as > > static struct smc_lo_systemeid LO_SYSTEM_EID = { > .seid_string = "SMC-SYSZ-LOSEID000000000", > .serial_number = "1000", > .type = "1000", > }; > > Is there anything else I need to pay attention to? > If you just want to use V2, such defination looks good. e.g. you can use some unique information from "lshw" > > 2) Seems only the first added smcd device determinate the system > smc_ism_v2_capable? If two > different smcd devices respectively with v1-indicated and > v2-indicated system_eid, will > the order in which they are registered affects the result of > smc_ism_v2_capable ? > see (*1) > > # [Q-3]: > > In attachment patch, I define a special CHID (0xFFFF) for loopback > device, as a kind of > 'unassociated ISM CHID' that not associated with any IP (OSA or > HiperSockets) interfaces. > > What's your opinion about this? > It looks good to me > > # [Q-4]: > > In current Linux implementation, server will select the first > successfully initialized device > from the candidates as the final selected one in > smc_find_ism_v2_device_serv(). > > for (i = 0; i < matches; i++) { > ini->smcd_version = SMC_V2; > ini->is_smcd = true; > ini->ism_selected = i; > rc = smc_listen_ism_init(new_smc, ini); > if (rc) { > smc_find_ism_store_rc(rc, ini); > /* try next active ISM device */ > continue; > } > return; /* matching and usable V2 ISM device found */ > } > > IMHO, maybe candidate devices should have different priorities? For > example, the loopback device > may be preferred to use if loopback is available. > IMO, I'd prefer such a order: ISM -> loopback -> RoCE Because ISM for SMC-D is our standard user case, not loopback. > > Best Regards, > Wen Gu > >>>>> >>>>> [1] >>>>> https://lore.kernel.org/netdev/20220720170048.20806-1-tonylu@linux.alibaba.com/ >>>>> [2] >>>>> https://lore.kernel.org/netdev/35d14144-28f7-6129-d6d3-ba16dae7a646@linux.ibm.com/ >>>>> [3] https://github.com/goldsborough/ipc-bench >>>>> >>>>> v1->v2 >>>>> 1. Fix some build WARNINGs complained by kernel test rebot >>>>> Reported-by: kernel test robot <lkp@intel.com> >>>>> 2. Add iperf3 test data. >>>>> >>>>> Wen Gu (5): >>>>> net/smc: introduce SMC-D loopback device >>>>> net/smc: choose loopback device in SMC-D communication >>>>> net/smc: add dmb attach and detach interface >>>>> net/smc: avoid data copy from sndbuf to peer RMB in SMC-D loopback >>>>> net/smc: logic of cursors update in SMC-D loopback connections >>>>> >>>>> include/net/smc.h | 3 + >>>>> net/smc/Makefile | 2 +- >>>>> net/smc/af_smc.c | 88 +++++++++++- >>>>> net/smc/smc_cdc.c | 59 ++++++-- >>>>> net/smc/smc_cdc.h | 1 + >>>>> net/smc/smc_clc.c | 4 +- >>>>> net/smc/smc_core.c | 62 +++++++++ >>>>> net/smc/smc_core.h | 2 + >>>>> net/smc/smc_ism.c | 39 +++++- >>>>> net/smc/smc_ism.h | 2 + >>>>> net/smc/smc_loopback.c | 358 >>>>> +++++++++++++++++++++++++++++++++++++++++++++++++ >>>>> net/smc/smc_loopback.h | 63 +++++++++ >>>>> 12 files changed, 662 insertions(+), 21 deletions(-) >>>>> create mode 100644 net/smc/smc_loopback.c >>>>> create mode 100644 net/smc/smc_loopback.h >>>>>
On 2023/1/16 19:01, Wenjia Zhang wrote: > > > On 12.01.23 13:12, Wen Gu wrote: >> >> >> On 2023/1/5 00:09, Alexandra Winter wrote: >>> >>> >>> On 21.12.22 14:14, Wen Gu wrote: >>>> >>>> >>>> On 2022/12/20 22:02, Niklas Schnelle wrote: >>>> >>>>> On Tue, 2022-12-20 at 11:21 +0800, Wen Gu wrote: >>>>>> Hi, all >>>>>> >>>>>> # Background >>>>>> >>>>>> As previously mentioned in [1], we (Alibaba Cloud) are trying to use SMC >>>>>> to accelerate TCP applications in cloud environment, improving inter-host >>>>>> or inter-VM communication. >>>>>> >>>>>> In addition of these, we also found the value of SMC-D in scenario of local >>>>>> inter-process communication, such as accelerate communication between containers >>>>>> within the same host. So this RFC tries to provide a SMC-D loopback solution >>>>>> in such scenario, to bring a significant improvement in latency and throughput >>>>>> compared to TCP loopback. >>>>>> >>>>>> # Design >>>>>> >>>>>> This patch set provides a kind of SMC-D loopback solution. >>>>>> >>>>>> Patch #1/5 and #2/5 provide an SMC-D based dummy device, preparing for the >>>>>> inter-process communication acceleration. Except for loopback acceleration, >>>>>> the dummy device can also meet the requirements mentioned in [2], which is >>>>>> providing a way to test SMC-D logic for broad community without ISM device. >>>>>> >>>>>> +------------------------------------------+ >>>>>> | +-----------+ +-----------+ | >>>>>> | | process A | | process B | | >>>>>> | +-----------+ +-----------+ | >>>>>> | ^ ^ | >>>>>> | | +---------------+ | | >>>>>> | | | SMC stack | | | >>>>>> | +--->| +-----------+ |<--| | >>>>>> | | | dummy | | | >>>>>> | | | device | | | >>>>>> | +-+-----------+-+ | >>>>>> | VM | >>>>>> +------------------------------------------+ >>>>>> >>>>>> Patch #3/5, #4/5, #5/5 provides a way to avoid data copy from sndbuf to RMB >>>>>> and improve SMC-D loopback performance. Through extending smcd_ops with two >>>>>> new semantic: attach_dmb and detach_dmb, sender's sndbuf shares the same >>>>>> physical memory region with receiver's RMB. The data copied from userspace >>>>>> to sender's sndbuf directly reaches the receiver's RMB without unnecessary >>>>>> memory copy in the same kernel. >>>>>> >>>>>> +----------+ +----------+ >>>>>> | socket A | | socket B | >>>>>> +----------+ +----------+ >>>>>> | ^ >>>>>> | +---------+ | >>>>>> regard as | | ----------| >>>>>> local sndbuf | B's | regard as >>>>>> | | RMB | local RMB >>>>>> |-------> | | >>>>>> +---------+ >>>>> >>>>> Hi Wen Gu, >>>>> >>>>> I maintain the s390 specific PCI support in Linux and would like to >>>>> provide a bit of background on this. You're surely wondering why we >>>>> even have a copy in there for our ISM virtual PCI device. To understand >>>>> why this copy operation exists and why we need to keep it working, one >>>>> needs a bit of s390 aka mainframe background. >>>>> >>>>> On s390 all (currently supported) native machines have a mandatory >>>>> machine level hypervisor. All OSs whether z/OS or Linux run either on >>>>> this machine level hypervisor as so called Logical Partitions (LPARs) >>>>> or as second/third/… level guests on e.g. a KVM or z/VM hypervisor that >>>>> in turn runs in an LPAR. Now, in terms of memory this machine level >>>>> hypervisor sometimes called PR/SM unlike KVM, z/VM, or VMWare is a >>>>> partitioning hypervisor without paging. This is one of the main reasons >>>>> for the very-near-native performance of the machine hypervisor as the >>>>> memory of its guests acts just like native RAM on other systems. It is >>>>> never paged out and always accessible to IOMMU translated DMA from >>>>> devices without the need for pinning pages and besides a trivial >>>>> offset/limit adjustment an LPAR's MMU does the same amount of work as >>>>> an MMU on a bare metal x86_64/ARM64 box. >>>>> >>>>> It also means however that when SMC-D is used to communicate between >>>>> LPARs via an ISM device there is no way of mapping the DMBs to the >>>>> same physical memory as there exists no MMU-like layer spanning >>>>> partitions that could do such a mapping. Meanwhile for machine level >>>>> firmware including the ISM virtual PCI device it is still possible to >>>>> _copy_ memory between different memory partitions. So yeah while I do >>>>> see the appeal of skipping the memcpy() for loopback or even between >>>>> guests of a paging hypervisor such as KVM, which can map the DMBs on >>>>> the same physical memory, we must keep in mind this original use case >>>>> requiring a copy operation. >>>>> >>>>> Thanks, >>>>> Niklas >>>>> >>>> >>>> Hi Niklas, >>>> >>>> Thank you so much for the complete and detailed explanation! This provides >>>> me a brand new perspective of s390 device that we hadn't dabbled in before. >>>> Now I understand why shared memory is unavailable between different LPARs. >>>> >>>> Our original intention of proposing loopback device and the incoming device >>>> (virtio-ism) for inter-VM is to use SMC-D to accelerate communication in the >>>> case with no existing s390 ISM devices. In our conception, s390 ISM device, >>>> loopback device and virtio-ism device are parallel and are abstracted by smcd_ops. >>>> >>>> +------------------------+ >>>> | SMC-D | >>>> +------------------------+ >>>> -------- smcd_ops --------- >>>> +------+ +------+ +------+ >>>> | s390 | | loop | |virtio| >>>> | ISM | | back | | -ism | >>>> | dev | | dev | | dev | >>>> +------+ +------+ +------+ >>>> >>>> We also believe that keeping the existing design and behavior of s390 ISM >>>> device is unshaken. What we want to get support for is some smcd_ops extension >>>> for devices with optional beneficial capability, such as nocopy here (Let's call >>>> it this for now), which is really helpful for us in inter-process and inter-VM >>>> scenario. >>>> >>>> And coincided with IBM's intention to add APIs between SMC-D and devices to >>>> support various devices for SMC-D, as mentioned in [2], we send out this RFC and >>>> the incoming virio-ism RFC, to provide some examples. >>>> >>>>>> >>>>>> # Benchmark Test >>>>>> >>>>>> * Test environments: >>>>>> - VM with Intel Xeon Platinum 8 core 2.50GHz, 16 GiB mem. >>>>>> - SMC sndbuf/RMB size 1MB. >>>>>> >>>>>> * Test object: >>>>>> - TCP: run on TCP loopback. >>>>>> - domain: run on UNIX domain. >>>>>> - SMC lo: run on SMC loopback device with patch #1/5 ~ #2/5. >>>>>> - SMC lo-nocpy: run on SMC loopback device with patch #1/5 ~ #5/5. >>>>>> >>>>>> 1. ipc-benchmark (see [3]) >>>>>> >>>>>> - ./<foo> -c 1000000 -s 100 >>>>>> >>>>>> TCP domain SMC-lo SMC-lo-nocpy >>>>>> Message >>>>>> rate (msg/s) 75140 129548(+72.41) 152266(+102.64%) 151914(+102.17%) >>>>> >>>>> Interesting that it does beat UNIX domain sockets. Also, see my below >>>>> comment for nginx/wrk as this seems very similar. >>>>> >>>>>> >>>>>> 2. sockperf >>>>>> >>>>>> - serv: <smc_run> taskset -c <cpu> sockperf sr --tcp >>>>>> - clnt: <smc_run> taskset -c <cpu> sockperf { tp | pp } --tcp --msg-size={ 64000 for tp | 14 for pp } -i >>>>>> 127.0.0.1 -t 30 >>>>>> >>>>>> TCP SMC-lo SMC-lo-nocpy >>>>>> Bandwidth(MBps) 4943.359 4936.096(-0.15%) 8239.624(+66.68%) >>>>>> Latency(us) 6.372 3.359(-47.28%) 3.25(-49.00%) >>>>>> >>>>>> 3. iperf3 >>>>>> >>>>>> - serv: <smc_run> taskset -c <cpu> iperf3 -s >>>>>> - clnt: <smc_run> taskset -c <cpu> iperf3 -c 127.0.0.1 -t 15 >>>>>> >>>>>> TCP SMC-lo SMC-lo-nocpy >>>>>> Bitrate(Gb/s) 40.5 41.4(+2.22%) 76.4(+88.64%) >>>>>> >>>>>> 4. nginx/wrk >>>>>> >>>>>> - serv: <smc_run> nginx >>>>>> - clnt: <smc_run> wrk -t 8 -c 500 -d 30 http://127.0.0.1:80 >>>>>> >>>>>> TCP SMC-lo SMC-lo-nocpy >>>>>> Requests/s 154643.22 220894.03(+42.84%) 226754.3(+46.63%) >>>>> >>>>> >>>>> This result is very interesting indeed. So with the much more realistic >>>>> nginx/wrk workload it seems to copy hurts much less than the >>>>> iperf3/sockperf would suggest while SMC-D itself seems to help more. >>>>> I'd hope that this translates to actual applications as well. Maybe >>>>> this makes SMC-D based loopback interesting even while keeping the >>>>> copy, at least until we can come up with a sane way to work a no-copy >>>>> variant into SMC-D? >>>>> >>>> >>>> I agree, nginx/wrk workload is much more realistic for many applications. >>>> >>>> But we also encounter many other cases similar to sockperf on the cloud, which >>>> requires high throughput, such as AI training and big data. >>>> >>>> So avoidance of copying between DMBs can help these cases a lot :) >>>> >>>>>> >>>>>> >>>>>> # Discussion >>>>>> >>>>>> 1. API between SMC-D and ISM device >>>>>> >>>>>> As Jan mentioned in [2], IBM are working on placing an API between SMC-D >>>>>> and the ISM device for easier use of different "devices" for SMC-D. >>>>>> >>>>>> So, considering that the introduction of attach_dmb or detach_dmb can >>>>>> effectively avoid data copying from sndbuf to RMB and brings obvious >>>>>> throughput advantages in inter-VM or inter-process scenarios, can the >>>>>> attach/detach semantics be taken into consideration when designing the >>>>>> API to make it a standard ISM device behavior? >>>> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ >>>>> >>>>> Due to the reasons explained above this behavior can't be emulated by >>>>> ISM devices at least not when crossing partitions. Not sure if we can >>>>> still incorporate it in the API and allow for both copying and >>>>> remapping SMC-D like devices, it definitely needs careful consideration >>>>> and I think also a better understanding of the benefit for real world >>>>> workloads. >>>>> >>>> >>>> Here I am not rigorous. >>>> >>>> Nocopy shouldn't be a standard ISM device behavior indeed. Actually we hope it be a >>>> standard optional _SMC-D_ device behavior and defined by smcd_ops. >>>> >>>> For devices don't support these options, like ISM device on s390 architecture, >>>> .attach_dmb/.detach_dmb and other reasonable extensions (which will be proposed to >>>> discuss in incoming virtio-ism RFC) can be set to NULL or return invalid. And for >>>> devices do support, they may be used for improving performance in some cases. >>>> >>>> In addition, can I know more latest news about the API design? :) , like its scale, will >>>> it be a almost refactor of existing interface or incremental patching? and its object, >>>> will it be tailored for exact ISM behavior or to reserve some options for other devices, >>>> like nocopy here? From my understanding of [2], it might be the latter? >>>> >>>>>> >>>>>> Maybe our RFC of SMC-D based inter-process acceleration (this one) and >>>>>> inter-VM acceleration (will coming soon, which is the update of [1]) >>>>>> can provide some examples for new API design. And we are very glad to >>>>>> discuss this on the mail list. >>>>>> >>>>>> 2. Way to select different ISM-like devices >>>>>> >>>>>> With the proposal of SMC-D loopback 'device' (this RFC) and incoming >>>>>> device used for inter-VM acceleration as update of [1], SMC-D has more >>>>>> options to choose from. So we need to consider that how to indicate >>>>>> supported devices, how to determine which one to use, and their priority... >>>>> >>>>> Agree on this part, though it is for the SMC maintainers to decide, I >>>>> think we would definitely want to be able to use any upcoming inter-VM >>>>> devices on s390 possibly also in conjunction with ISM devices for >>>>> communication across partitions. >>>>> >>>> >>>> Yes, this part needs to be discussed with SMC maintainers. And thank you, we are very glad >>>> if our devices can be applied on s390 through the efforts. >>>> >>>> >>>> Best Regards, >>>> Wen Gu >>>> >>>>>> >>>>>> IMHO, this may require an update of CLC message and negotiation mechanism. >>>>>> Again, we are very glad to discuss this with you on the mailing list. >>> >>> As described in >>> SMC protocol (including SMC-D): >>> https://www.ibm.com/support/pages/system/files/inline-files/IBM%20Shared%20Memory%20Communications%20Version%202_2.pdf >>> the CLC messages provide a list of up to 8 ISM devices to chose from. >>> So I would hope that we can use the existing protocol. >>> >>> The challenge will be to define GID (Global Interface ID) and CHID (a fabric ID) in >>> a meaningful way for the new devices. >>> There is always smcd_ops->query_remote_gid() as a safety net. But the idea is that >>> a CHID mismatch is a fast way to tell that these 2 interfaces do match. >>> >>> >> > > FYI, we just sent the rest part of the API to the net-next > https://lore.kernel.org/netdev/20230116092712.10176-1-jaka@linux.ibm.com/T/#t, > which should answer some questions in your patch series. > Hi Wenjia, Thanks for notification and your excellent work! I will learn about the new API and find if it fits our needs. If all goes well I will refactor my RFC based on the new API and send a v2, most likely after the Chinese New Year. > >> Hi Winter and all, >> >> Thanks for your reply and suggestions! And sorry for my late reply because it took me >> some time to understand SMC-Dv2 protocol and implementation. >> >> I agree with your opinion. The existing SMC-Dv2 protocol whose CLC messages include >> ism_dev[] list can solve the devices negotiation problem. And I am very willing to use >> the existing protocol, because we all know that the protocol update is a long and complex >> process. >> >> If I understand correctly, SMC-D loopback(dummy) device can coordinate with existing >> SMC-Dv2 protocol as follows. If there is any mistake, please point out. >> >> >> # Initialization >> >> - Initialize the loopback device with unique GID [Q-1]. >> >> - Register the loopback device as SMC-Dv2-capable device with a system_eid whose 24th >> or 28th byte is non-zero [Q-2], so that this system's smc_ism_v2_capable will be set >> to TRUE and SMC-Dv2 is available. >> > The decision point is the VLAN_ID, if it is x1FFF, the device will support V2. i.e. If you can have subnet with VLAN_ID > x1FFF, then the SEID is necessary, so that the series or types is non-zero. (*1) In case there is any misunderstanding between us, I would like to rephrase my [Q-2] question: int smcd_register_dev(struct smcd_dev *smcd) { <...> mutex_lock(&smcd_dev_list.mutex); if (list_empty(&smcd_dev_list.list)) { u8 *system_eid = NULL; smcd->ops->get_system_eid(smcd, &system_eid); if (system_eid[24] != '0' || system_eid[28] != '0') { smc_ism_v2_capable = true; memcpy(smc_ism_v2_system_eid, system_eid, SMC_MAX_EID_LEN); } } <...> } It can be inferred from smcd_register_dev() that: 1) The 24th and 28th byte are special and determinate whether smc_ism_v2_capable is true. Besides these, do other bytes of system_eid have hidden meanings that need attention ? 2) Only when smcd_dev_list is empty, the added smcd_dev will be checked, and its system_eid determinates whether smc_ism_v2_capable is true. Why only the first added device will be checked ? If the first added smcd_dev has an system_eid whose 24th and 28th bytes are zero, and the second added smcd_dev has an system_eid whose 24th and 28th bytes are non-zero. Should smc_ism_v2_capable be true, since the second smcd_dev has v2-indicated system_eid ? >> >> # Proposal >> >> - Find the loopback device from the smcd_dev_list in smc_find_ism_v2_device_clnt(); >> >> - Record the SEID, GID and CHID[Q-3] of loopback device in the v2 extension part of CLC >> proposal message. >> >> >> # Accept >> >> - Check the GID/CHID list and SEID in CLC proposal message, and find local matched ISM >> device from smcd_dev_list in smc_find_ism_v2_device_serv(). If both sides of the >> communication are in the same VM and share the same loopback device, the SEID, GID and >> CHID will match and loopback device will be chosen [Q-4]. >> >> - Record the loopback device's GID/CHID and matched SEID into CLC accept message. >> >> >> # Confirm >> >> - Confirm the server-selected device (loopback device) accordingto CLC accept messages. >> >> - Record the loopback device's GID/CHID and server-selected SEID in CLC confirm message. >> >> >> Follow the above process, I supplement a patch based on this RFC in the email attachment. >> With the attachment patch, SMC-D loopback will switch to use SMC-Dv2 protocol. >> >> >> >> And in the above process, there are something I want to consult and discuss, which is marked >> with '[Q-*]' in the above description. >> >> # [Q-1]: >> >> The GID of loopback device is randomly generated in this RFC patch set, but I will find a way >> to unique the GID in formal patches. Any suggestions are welcome. >> > I think the randowmly generated GID is fine in your case, which is equivalent to the IP address. Since whether the two sides can communicate through the loopback will be judged by whether the gid of their loopback device is equal, the random GID may bring the risk of misjudgment because it may not be unique. But considering this is an RFC, I simply used random GIDs. >> >> # [Q-2]: >> >> In Linux implementation, the system_eid of the first registered smcd device will determinate >> system's smc_ism_v2_capable (see smcd_register_dev()). >> >> And I wonder that >> >> 1) How to define the system_eid? It can be inferred from the code that the 24th and 28th byte >> are special for SMC-Dv2. So in attachment patch, I define the loopback device SEID as >> >> static struct smc_lo_systemeid LO_SYSTEM_EID = { >> .seid_string = "SMC-SYSZ-LOSEID000000000", >> .serial_number = "1000", >> .type = "1000", >> }; >> >> Is there anything else I need to pay attention to? >> > If you just want to use V2, such defination looks good. > e.g. you can use some unique information from "lshw" OK, thank you. >> >> 2) Seems only the first added smcd device determinate the system smc_ism_v2_capable? If two >> different smcd devices respectively with v1-indicated and v2-indicated system_eid, will >> the order in which they are registered affects the result of smc_ism_v2_capable ? >> > see (*1) >> >> # [Q-3]: >> >> In attachment patch, I define a special CHID (0xFFFF) for loopback device, as a kind of >> 'unassociated ISM CHID' that not associated with any IP (OSA or HiperSockets) interfaces. >> >> What's your opinion about this? >> > It looks good to me OK. >> >> # [Q-4]: >> >> In current Linux implementation, server will select the first successfully initialized device >> from the candidates as the final selected one in smc_find_ism_v2_device_serv(). >> >> for (i = 0; i < matches; i++) { >> ini->smcd_version = SMC_V2; >> ini->is_smcd = true; >> ini->ism_selected = i; >> rc = smc_listen_ism_init(new_smc, ini); >> if (rc) { >> smc_find_ism_store_rc(rc, ini); >> /* try next active ISM device */ >> continue; >> } >> return; /* matching and usable V2 ISM device found */ >> } >> >> IMHO, maybe candidate devices should have different priorities? For example, the loopback device >> may be preferred to use if loopback is available. >> > IMO, I'd prefer such a order: ISM -> loopback -> RoCE > Because ISM for SMC-D is our standard user case, not loopback. OK, will follow this order. >> >> Best Regards, >> Wen Gu >> >>>>>> >>>>>> [1] https://lore.kernel.org/netdev/20220720170048.20806-1-tonylu@linux.alibaba.com/ >>>>>> [2] https://lore.kernel.org/netdev/35d14144-28f7-6129-d6d3-ba16dae7a646@linux.ibm.com/ >>>>>> [3] https://github.com/goldsborough/ipc-bench >>>>>> >>>>>> v1->v2 >>>>>> 1. Fix some build WARNINGs complained by kernel test rebot >>>>>> Reported-by: kernel test robot <lkp@intel.com> >>>>>> 2. Add iperf3 test data. >>>>>> >>>>>> Wen Gu (5): >>>>>> net/smc: introduce SMC-D loopback device >>>>>> net/smc: choose loopback device in SMC-D communication >>>>>> net/smc: add dmb attach and detach interface >>>>>> net/smc: avoid data copy from sndbuf to peer RMB in SMC-D loopback >>>>>> net/smc: logic of cursors update in SMC-D loopback connections >>>>>> >>>>>> include/net/smc.h | 3 + >>>>>> net/smc/Makefile | 2 +- >>>>>> net/smc/af_smc.c | 88 +++++++++++- >>>>>> net/smc/smc_cdc.c | 59 ++++++-- >>>>>> net/smc/smc_cdc.h | 1 + >>>>>> net/smc/smc_clc.c | 4 +- >>>>>> net/smc/smc_core.c | 62 +++++++++ >>>>>> net/smc/smc_core.h | 2 + >>>>>> net/smc/smc_ism.c | 39 +++++- >>>>>> net/smc/smc_ism.h | 2 + >>>>>> net/smc/smc_loopback.c | 358 +++++++++++++++++++++++++++++++++++++++++++++++++ >>>>>> net/smc/smc_loopback.h | 63 +++++++++ >>>>>> 12 files changed, 662 insertions(+), 21 deletions(-) >>>>>> create mode 100644 net/smc/smc_loopback.c >>>>>> create mode 100644 net/smc/smc_loopback.h >>>>>>
On 18.01.23 13:15, Wen Gu wrote: > > > On 2023/1/16 19:01, Wenjia Zhang wrote: >> >> >> On 12.01.23 13:12, Wen Gu wrote: >>> >>> >>> On 2023/1/5 00:09, Alexandra Winter wrote: >>>> >>>> >>>> On 21.12.22 14:14, Wen Gu wrote: >>>>> >>>>> >>>>> On 2022/12/20 22:02, Niklas Schnelle wrote: >>>>> >>>>>> On Tue, 2022-12-20 at 11:21 +0800, Wen Gu wrote: [...] >>>>>>> >>>>>>> 2. Way to select different ISM-like devices >>>>>>> >>>>>>> With the proposal of SMC-D loopback 'device' (this RFC) and incoming >>>>>>> device used for inter-VM acceleration as update of [1], SMC-D has more >>>>>>> options to choose from. So we need to consider that how to indicate >>>>>>> supported devices, how to determine which one to use, and their priority... >>>>>> [...] >>>>>>> >>>>>>> IMHO, this may require an update of CLC message and negotiation mechanism. >>>>>>> Again, we are very glad to discuss this with you on the mailing list. >>>> >>>> As described in >>>> SMC protocol (including SMC-D): https://www.ibm.com/support/pages/system/files/inline-files/IBM%20Shared%20Memory%20Communications%20Version%202_2.pdf >>>> the CLC messages provide a list of up to 8 ISM devices to chose from. >>>> So I would hope that we can use the existing protocol. >>>> >>>> The challenge will be to define GID (Global Interface ID) and CHID (a fabric ID) in >>>> a meaningful way for the new devices. [...] >>> >>> I agree with your opinion. The existing SMC-Dv2 protocol whose CLC messages include >>> ism_dev[] list can solve the devices negotiation problem. And I am very willing to use >>> the existing protocol, because we all know that the protocol update is a long and complex >>> process. >>> >>> If I understand correctly, SMC-D loopback(dummy) device can coordinate with existing >>> SMC-Dv2 protocol as follows. If there is any mistake, please point out. >>> >>> >>> # Initialization >>> >>> - Initialize the loopback device with unique GID [Q-1]. As you point out below [Q-1], the issue here is the uniqueness of the 8 byte GID across all possible SMC peers (they could all offer loopback as a choice). I was pondering some more about this and I am wondering, whether there is another way to detect that source and target of the CLC proposal are the same Linux instance (without using the content of the CLC message). One idea could be: If the netdev used for the CLC proposal is a local interface, then use smcd-loopback. What do you think? Let me try to summarize my pondering as a base for future discussion: SMC CLC handshake as described in SMC protocol: ------------------------------------------------ The purpose of the handshake is to efficiently find out whether and how SMC can be used to communicate with a TCP peer (happens over a standard netdevice). For this purpose the following fields are exchanged: + UEID – (32 byte) Defined per system (OS instance). User defined group of systems that should use SMC between each other. If the peer has only different UEIDs: Don’t try to use SMC with this peer. + SEID – (32 byte) Defined per system. Maximum space of systems that are able to use SMC-D between each other. Equals to the machine hardware / first level hypervisor. (in s390 KVM is at least the second level hipervisor) Is unique per machine, derived from unique machine ID. If the peer has a different SEIDs: Don’t try to use SMC-D with this peer. + CHID – (2 Bytes) Defined per SMC-D interface. Fabric ID of this interface. Unique against other fabrics on this machine. Try to find a pair of SMC-D interfaces on the same fabric for the 2 peers to use for this SMC-D connection. + GID – (8 bytes) Defined per SMC-D interface. ID of this interface. For s390 ISM is actually globally unique. But for SMC-D protocol purpose just needs to be unique within the CHID (fabric). Use it to identify the chosen interfaces of the 2 peers. Important usecases: As the CLC handshake is exchanged via standard netdevice, the 2 peers can be: - on different machines - on same machine but different first level guest - on same KVM host but different KVM guest, - on same KVM guest (loopback case) and the proposal list can include any combination of: - one or more RDMA devices (to use with SMC-R) - one or more s390 ISM devices on one or more CHIDs - future: virtio-smcd interface(s) So for loopback there are 2 options to think about, that can be implemented with today’s SMC-D protocol definition: (1) Smc-d Loopback is listed in the SMC-D proposal list Loopback could be one interface in the SMC-Dv2 list of up to 8 CHID/GID pairs proposed. We could use CHID 0xFFFF to point out this is a loopback. And then only use this GID, if it belongs to both peers. (this is may be a small add-on to today's protocol) CON: This requires a unique loopback-GID for every OS instance (on this SEID). Must also be unique against any other OS that would ever implement SMC-D loopback, because this could be a handshake with any OS on this SEID. PRO: It works in all cases where SMC works today. (2)Find out that both peers of the CLC handshake are actually the same Linux instance, without using the values of the proposal One idea is that if a local netdev is used for the handshake, then we are in the same instance. CON: I guess this should work, but may not cover all usecases, where smcd-loopback could be desired. (namespaces, etc..) I still have some hope that there is another way to detect this somehow… ideas are very welcome. PRO: This is independent of the SMC protocol. It would be a Linux-only solution, actually even *this* Linux only, future implementations could differ. >>> >>> - Register the loopback device as SMC-Dv2-capable device with a system_eid whose 24th >>> or 28th byte is non-zero [Q-2], so that this system's smc_ism_v2_capable will be set >>> to TRUE and SMC-Dv2 is available. >>> >> The decision point is the VLAN_ID, if it is x1FFF, the device will support V2. i.e. If you can have subnet with VLAN_ID x1FFF, then the SEID is necessary, so that the series or types is non-zero. (*1) > I guess there is some misunderstanding of today's code. The invalid VLAN_ID 0x1FFF is used to signal to s390 ISM hardware that SMC-Dv2 is used, where the CLC handshake does not have to be between peers on the same IP subnet. If we cannot set x1FFF, then this is old hardware and we have to use SMC-Dv1 on this machine. (all ISM interfaces are same hardware level). If we can use SMC-Dv2 then we need to determine the SEID of this machine, so we can use it in the CLC handshake. (if we run on v1 hardware, there is no need to determine the SEID) > In case there is any misunderstanding between us, I would like to rephrase my [Q-2] question: > > int smcd_register_dev(struct smcd_dev *smcd) > { > <...> > mutex_lock(&smcd_dev_list.mutex); > if (list_empty(&smcd_dev_list.list)) { > u8 *system_eid = NULL; > > smcd->ops->get_system_eid(smcd, &system_eid); > if (system_eid[24] != '0' || system_eid[28] != '0') { > smc_ism_v2_capable = true; > memcpy(smc_ism_v2_system_eid, system_eid, > SMC_MAX_EID_LEN); > } > } > <...> > } > > It can be inferred from smcd_register_dev() that: > > 1) The 24th and 28th byte are special and determinate whether smc_ism_v2_capable is true. > Besides these, do other bytes of system_eid have hidden meanings that need attention ? > > 2) Only when smcd_dev_list is empty, the added smcd_dev will be checked, and its system_eid > determinates whether smc_ism_v2_capable is true. Why only the first added device will be > checked ? > > If the first added smcd_dev has an system_eid whose 24th and 28th bytes are zero, and the > second added smcd_dev has an system_eid whose 24th and 28th bytes are non-zero. Should > smc_ism_v2_capable be true, since the second smcd_dev has v2-indicated system_eid ? > This was a rather indirect way to determine smc_ism_v2_capable, which is improved in the ism patches currently under review. >>> >>> # Proposal >>> >>> - Find the loopback device from the smcd_dev_list in smc_find_ism_v2_device_clnt(); >>> >>> - Record the SEID, GID and CHID[Q-3] of loopback device in the v2 extension part of CLC >>> proposal message. >>> >>> >>> # Accept >>> >>> - Check the GID/CHID list and SEID in CLC proposal message, and find local matched ISM >>> device from smcd_dev_list in smc_find_ism_v2_device_serv(). If both sides of the >>> communication are in the same VM and share the same loopback device, the SEID, GID and >>> CHID will match and loopback device will be chosen [Q-4]. >>> >>> - Record the loopback device's GID/CHID and matched SEID into CLC accept message. >>> >>> >>> # Confirm >>> >>> - Confirm the server-selected device (loopback device) accordingto CLC accept messages. >>> >>> - Record the loopback device's GID/CHID and server-selected SEID in CLC confirm message. >>> >>> >>> Follow the above process, I supplement a patch based on this RFC in the email attachment. >>> With the attachment patch, SMC-D loopback will switch to use SMC-Dv2 protocol. >>> >>> >>> >>> And in the above process, there are something I want to consult and discuss, which is marked >>> with '[Q-*]' in the above description. >>> >>> # [Q-1]: >>> >>> The GID of loopback device is randomly generated in this RFC patch set, but I will find a way >>> to unique the GID in formal patches. Any suggestions are welcome. >>> >> I think the randowmly generated GID is fine in your case, which is equivalent to the IP address. > > Since whether the two sides can communicate through the loopback will be judged by whether the > gid of their loopback device is equal, the random GID may bring the risk of misjudgment because > it may not be unique. But considering this is an RFC, I simply used random GIDs. I share your concerns about using random 8 byte numbers for all possible instances. Collisions may be unlikely, but if they happen, they cannot be detected and have nasty effects. > >>> >>> # [Q-2]: >>> >>> In Linux implementation, the system_eid of the first registered smcd device will determinate >>> system's smc_ism_v2_capable (see smcd_register_dev()). See above, this is a rather indirect correlation. >>> >>> And I wonder that >>> >>> 1) How to define the system_eid? It can be inferred from the code that the 24th and 28th byte >>> are special for SMC-Dv2. So in attachment patch, I define the loopback device SEID as >>> >>> static struct smc_lo_systemeid LO_SYSTEM_EID = { >>> .seid_string = "SMC-SYSZ-LOSEID000000000", >>> .serial_number = "1000", >>> .type = "1000", >>> }; >>> >>> Is there anything else I need to pay attention to? >>> >> If you just want to use V2, such defination looks good. >> e.g. you can use some unique information from "lshw" > > OK, thank you. > As mentioned above: + SEID – (32 byte) Defined per system. Maximum space of systems that are able to use SMC-D between each other. Equals to the machine hardware / first level hypervisor. (in s390 KVM is at least the second level hipervisor) Is unique per machine, derived from unique machine ID. If the peer has a different SEIDs: Don’t try to use SMC-D with this peer. We need to continue to use today's values on s390 architecture for backward compatibility! Other architectures need to also use values that uniquely identify the machine it is running on. >>> >>> 2) Seems only the first added smcd device determinate the system smc_ism_v2_capable? If two >>> different smcd devices respectively with v1-indicated and v2-indicated system_eid, will >>> the order in which they are registered affects the result of smc_ism_v2_capable ? >>> >> see (*1) see above: all s390 ISM interfaces on a machine are same hardware level. >>> >>> # [Q-3]: >>> >>> In attachment patch, I define a special CHID (0xFFFF) for loopback device, as a kind of >>> 'unassociated ISM CHID' that not associated with any IP (OSA or HiperSockets) interfaces. >>> >>> What's your opinion about this? >>> >> It looks good to me > > OK. This maybe a small add-on to today's protocol, as a special CHID number that is evaluated differently. But IMHO it would fit with the purpose of the VHCID/GID pairs. 0xFFFF cannot appear on today's s390ISM CHIDs. So this could be backwards compatible. > >>> >>> # [Q-4]: >>> >>> In current Linux implementation, server will select the first successfully initialized device >>> from the candidates as the final selected one in smc_find_ism_v2_device_serv(). >>> >>> for (i = 0; i < matches; i++) { >>> ini->smcd_version = SMC_V2; >>> ini->is_smcd = true; >>> ini->ism_selected = i; >>> rc = smc_listen_ism_init(new_smc, ini); >>> if (rc) { >>> smc_find_ism_store_rc(rc, ini); >>> /* try next active ISM device */ >>> continue; >>> } >>> return; /* matching and usable V2 ISM device found */ >>> } >>> >>> IMHO, maybe candidate devices should have different priorities? For example, the loopback device >>> may be preferred to use if loopback is available. >>> >> IMO, I'd prefer such a order: ISM -> loopback -> RoCE >> Because ISM for SMC-D is our standard user case, not loopback. > > OK, will follow this order. My initial thought would be: loopback -> ISM -> RoCE, just as it is for netdev loopback. But that only makes sense if loopback performs better than ISM. Can we postpone that decision until we have measurements?
On 2023/1/19 20:30, Alexandra Winter wrote: > > > On 18.01.23 13:15, Wen Gu wrote: >> >> >> On 2023/1/16 19:01, Wenjia Zhang wrote: >>> >>> >>> On 12.01.23 13:12, Wen Gu wrote: >>>> >>>> >>>> On 2023/1/5 00:09, Alexandra Winter wrote: >>>>> >>>>> >>>>> On 21.12.22 14:14, Wen Gu wrote: >>>>>> >>>>>> >>>>>> On 2022/12/20 22:02, Niklas Schnelle wrote: >>>>>> >>>>>>> On Tue, 2022-12-20 at 11:21 +0800, Wen Gu wrote: > [...] >>>>>>>> >>>>>>>> 2. Way to select different ISM-like devices >>>>>>>> >>>>>>>> With the proposal of SMC-D loopback 'device' (this RFC) and incoming >>>>>>>> device used for inter-VM acceleration as update of [1], SMC-D has more >>>>>>>> options to choose from. So we need to consider that how to indicate >>>>>>>> supported devices, how to determine which one to use, and their priority... >>>>>>> > [...] >>>>>>>> >>>>>>>> IMHO, this may require an update of CLC message and negotiation mechanism. >>>>>>>> Again, we are very glad to discuss this with you on the mailing list. >>>>> >>>>> As described in >>>>> SMC protocol (including SMC-D): https://www.ibm.com/support/pages/system/files/inline-files/IBM%20Shared%20Memory%20Communications%20Version%202_2.pdf >>>>> the CLC messages provide a list of up to 8 ISM devices to chose from. >>>>> So I would hope that we can use the existing protocol. >>>>> >>>>> The challenge will be to define GID (Global Interface ID) and CHID (a fabric ID) in >>>>> a meaningful way for the new devices. > [...] >>>> >>>> I agree with your opinion. The existing SMC-Dv2 protocol whose CLC messages include >>>> ism_dev[] list can solve the devices negotiation problem. And I am very willing to use >>>> the existing protocol, because we all know that the protocol update is a long and complex >>>> process. >>>> >>>> If I understand correctly, SMC-D loopback(dummy) device can coordinate with existing >>>> SMC-Dv2 protocol as follows. If there is any mistake, please point out. >>>> >>>> >>>> # Initialization >>>> >>>> - Initialize the loopback device with unique GID [Q-1]. > > As you point out below [Q-1], the issue here is the uniqueness of the 8 byte GID > across all possible SMC peers (they could all offer loopback as a choice). > > I was pondering some more about this and I am wondering, whether there is another > way to detect that source and target of the CLC proposal are the same Linux instance > (without using the content of the CLC message). > One idea could be: > If the netdev used for the CLC proposal is a local interface, then use smcd-loopback. > What do you think? > Hi Winter, Thanks a lot for your suggestions, and forgive me for the delay in replying due to the vacation. My opinions are below. > Let me try to summarize my pondering as a base for future discussion: > > SMC CLC handshake as described in SMC protocol: > ------------------------------------------------ > The purpose of the handshake is to efficiently find out whether and how SMC can > be used to communicate with a TCP peer (happens over a standard netdevice). > For this purpose the following fields are exchanged: > > + UEID – (32 byte) Defined per system (OS instance). > User defined group of systems that should use SMC between each other. > If the peer has only different UEIDs: Don’t try to use SMC with this peer. > > + SEID – (32 byte) Defined per system. > Maximum space of systems that are able to use SMC-D between each other. > Equals to the machine hardware / first level hypervisor. > (in s390 KVM is at least the second level hipervisor) > Is unique per machine, derived from unique machine ID. > If the peer has a different SEIDs: Don’t try to use SMC-D with this peer. > > + CHID – (2 Bytes) Defined per SMC-D interface. > Fabric ID of this interface. Unique against other fabrics on this machine. > Try to find a pair of SMC-D interfaces on the same fabric for the 2 peers > to use for this SMC-D connection. > > + GID – (8 bytes) Defined per SMC-D interface. > ID of this interface. > For s390 ISM is actually globally unique. But for SMC-D protocol purpose > just needs to be unique within the CHID (fabric). > Use it to identify the chosen interfaces of the 2 peers. > > Important usecases: > As the CLC handshake is exchanged via standard netdevice, the 2 peers can be: > - on different machines > - on same machine but different first level guest > - on same KVM host but different KVM guest, > - on same KVM guest (loopback case) > > and the proposal list can include any combination of: > - one or more RDMA devices (to use with SMC-R) > - one or more s390 ISM devices on one or more CHIDs > - future: virtio-smcd interface(s) > > > So for loopback there are 2 options to think about, that can be implemented > with today’s SMC-D protocol definition: > > (1) Smc-d Loopback is listed in the SMC-D proposal list > > Loopback could be one interface in the SMC-Dv2 list of up to 8 CHID/GID pairs proposed. > We could use CHID 0xFFFF to point out this is a loopback. > And then only use this GID, if it belongs to both peers. (this is may be a small > add-on to today's protocol) > CON: This requires a unique loopback-GID for every OS instance (on this SEID). > Must also be unique against any other OS that would ever implement SMC-D > loopback, because this could be a handshake with any OS on this SEID. > PRO: It works in all cases where SMC works today. > > (2)Find out that both peers of the CLC handshake are actually the same > Linux instance, without using the values of the proposal > > One idea is that if a local netdev is used for the handshake, then we are in > the same instance. > CON: I guess this should work, but may not cover all usecases, where > smcd-loopback could be desired. (namespaces, etc..) > I still have some hope that there is another way to detect this somehow… > ideas are very welcome. > PRO: This is independent of the SMC protocol. It would be a Linux-only solution, > actually even *this* Linux only, future implementations could differ. > I totally agree with your summary. I also considered option #2 at the beginning, like through established clcsock's fib to judge whether the CLC proposal target is local. But as you mentioned, it is hard to cover all possible usecases, a particular example is that two containers in the same VM communicate with each other through physical NICs. So IMHO, it seems hardly to avoid exchanging information (like GID) between each side to confirm both sides are on the same Linux instance. Please correct me if I'm wrong. For now, I perfer the option #1. But option #2 is still one of our options. All ideas are welcome. >>>> >>>> - Register the loopback device as SMC-Dv2-capable device with a system_eid whose 24th >>>> or 28th byte is non-zero [Q-2], so that this system's smc_ism_v2_capable will be set >>>> to TRUE and SMC-Dv2 is available. >>>> >>> The decision point is the VLAN_ID, if it is x1FFF, the device will support V2. i.e. If you can have subnet with VLAN_ID x1FFF, then the SEID is necessary, so that the series or types is non-zero. (*1) >> > > I guess there is some misunderstanding of today's code. > The invalid VLAN_ID 0x1FFF is used to signal to s390 ISM hardware that SMC-Dv2 is used, where > the CLC handshake does not have to be between peers on the same IP subnet. > If we cannot set x1FFF, then this is old hardware and we have to use SMC-Dv1 on this machine. > (all ISM interfaces are same hardware level). > If we can use SMC-Dv2 then we need to determine the SEID of this machine, so we can use it > in the CLC handshake. (if we run on v1 hardware, there is no need to determine the SEID) > Thanks for explanation, now I understand. >> In case there is any misunderstanding between us, I would like to rephrase my [Q-2] question: >> >> int smcd_register_dev(struct smcd_dev *smcd) >> { >> <...> >> mutex_lock(&smcd_dev_list.mutex); >> if (list_empty(&smcd_dev_list.list)) { >> u8 *system_eid = NULL; >> >> smcd->ops->get_system_eid(smcd, &system_eid); >> if (system_eid[24] != '0' || system_eid[28] != '0') { >> smc_ism_v2_capable = true; >> memcpy(smc_ism_v2_system_eid, system_eid, >> SMC_MAX_EID_LEN); >> } >> } >> <...> >> } >> >> It can be inferred from smcd_register_dev() that: >> >> 1) The 24th and 28th byte are special and determinate whether smc_ism_v2_capable is true. >> Besides these, do other bytes of system_eid have hidden meanings that need attention ? >> >> 2) Only when smcd_dev_list is empty, the added smcd_dev will be checked, and its system_eid >> determinates whether smc_ism_v2_capable is true. Why only the first added device will be >> checked ? >> >> If the first added smcd_dev has an system_eid whose 24th and 28th bytes are zero, and the >> second added smcd_dev has an system_eid whose 24th and 28th bytes are non-zero. Should >> smc_ism_v2_capable be true, since the second smcd_dev has v2-indicated system_eid ? >> > > > This was a rather indirect way to determine smc_ism_v2_capable, > which is improved in the ism patches currently under review. > OK, thanks. >>>> >>>> # Proposal >>>> >>>> - Find the loopback device from the smcd_dev_list in smc_find_ism_v2_device_clnt(); >>>> >>>> - Record the SEID, GID and CHID[Q-3] of loopback device in the v2 extension part of CLC >>>> proposal message. >>>> >>>> >>>> # Accept >>>> >>>> - Check the GID/CHID list and SEID in CLC proposal message, and find local matched ISM >>>> device from smcd_dev_list in smc_find_ism_v2_device_serv(). If both sides of the >>>> communication are in the same VM and share the same loopback device, the SEID, GID and >>>> CHID will match and loopback device will be chosen [Q-4]. >>>> >>>> - Record the loopback device's GID/CHID and matched SEID into CLC accept message. >>>> >>>> >>>> # Confirm >>>> >>>> - Confirm the server-selected device (loopback device) accordingto CLC accept messages. >>>> >>>> - Record the loopback device's GID/CHID and server-selected SEID in CLC confirm message. >>>> >>>> >>>> Follow the above process, I supplement a patch based on this RFC in the email attachment. >>>> With the attachment patch, SMC-D loopback will switch to use SMC-Dv2 protocol. >>>> >>>> >>>> >>>> And in the above process, there are something I want to consult and discuss, which is marked >>>> with '[Q-*]' in the above description. >>>> >>>> # [Q-1]: >>>> >>>> The GID of loopback device is randomly generated in this RFC patch set, but I will find a way >>>> to unique the GID in formal patches. Any suggestions are welcome. >>>> >>> I think the randowmly generated GID is fine in your case, which is equivalent to the IP address. >> >> Since whether the two sides can communicate through the loopback will be judged by whether the >> gid of their loopback device is equal, the random GID may bring the risk of misjudgment because >> it may not be unique. But considering this is an RFC, I simply used random GIDs. > > I share your concerns about using random 8 byte numbers for all possible instances. > Collisions may be unlikely, but if they happen, they cannot be detected and have nasty effects. > Yes, I'm thinking about it and trying to find a way to avoid collision, or safely fallback if collision happens. >> >>>> >>>> # [Q-2]: >>>> >>>> In Linux implementation, the system_eid of the first registered smcd device will determinate >>>> system's smc_ism_v2_capable (see smcd_register_dev()). > > See above, this is a rather indirect correlation. > >>>> >>>> And I wonder that >>>> >>>> 1) How to define the system_eid? It can be inferred from the code that the 24th and 28th byte >>>> are special for SMC-Dv2. So in attachment patch, I define the loopback device SEID as >>>> >>>> static struct smc_lo_systemeid LO_SYSTEM_EID = { >>>> .seid_string = "SMC-SYSZ-LOSEID000000000", >>>> .serial_number = "1000", >>>> .type = "1000", >>>> }; >>>> >>>> Is there anything else I need to pay attention to? >>>> >>> If you just want to use V2, such defination looks good. >>> e.g. you can use some unique information from "lshw" >> >> OK, thank you. >> > > As mentioned above: > + SEID – (32 byte) Defined per system. > Maximum space of systems that are able to use SMC-D between each other. > Equals to the machine hardware / first level hypervisor. > (in s390 KVM is at least the second level hipervisor) > Is unique per machine, derived from unique machine ID. > If the peer has a different SEIDs: Don’t try to use SMC-D with this peer. > > We need to continue to use today's values on s390 architecture for backward compatibility! > Other architectures need to also use values that uniquely identify the machine it is > running on. > Yes, I agree. IIUC, on s390 architecture, if machine's ISM hardware is V2 capable (can set VLAN_ID 0x1FFF), then SEID will be set according to unique machine ID (for s390 architecture, it will be the ID returned from get_cpu_id() in arch/s390/include/asm/processor.h). So, - On s390 architecture, SMC-D loopback device should use the same SEID as ISM devices on the same machine. If SMC-D loopback only supports V2 and machine is not V2 capable, then SMC-D loopback should not be used. - On the other architecture except s390, we need to find a similar way to generate a machine-unique ID as part of SEID. (perhaps need a unified helper to do so ?) >>>> >>>> 2) Seems only the first added smcd device determinate the system smc_ism_v2_capable? If two >>>> different smcd devices respectively with v1-indicated and v2-indicated system_eid, will >>>> the order in which they are registered affects the result of smc_ism_v2_capable ? >>>> >>> see (*1) > > see above: all s390 ISM interfaces on a machine are same hardware level. > OK. >>>> >>>> # [Q-3]: >>>> >>>> In attachment patch, I define a special CHID (0xFFFF) for loopback device, as a kind of >>>> 'unassociated ISM CHID' that not associated with any IP (OSA or HiperSockets) interfaces. >>>> >>>> What's your opinion about this? >>>> >>> It looks good to me >> >> OK. > > This maybe a small add-on to today's protocol, as a special CHID number that is evaluated > differently. But IMHO it would fit with the purpose of the VHCID/GID pairs. > 0xFFFF cannot appear on today's s390ISM CHIDs. So this could be backwards compatible. > Thanks, that's a good news. >> >>>> >>>> # [Q-4]: >>>> >>>> In current Linux implementation, server will select the first successfully initialized device >>>> from the candidates as the final selected one in smc_find_ism_v2_device_serv(). >>>> >>>> for (i = 0; i < matches; i++) { >>>> ini->smcd_version = SMC_V2; >>>> ini->is_smcd = true; >>>> ini->ism_selected = i; >>>> rc = smc_listen_ism_init(new_smc, ini); >>>> if (rc) { >>>> smc_find_ism_store_rc(rc, ini); >>>> /* try next active ISM device */ >>>> continue; >>>> } >>>> return; /* matching and usable V2 ISM device found */ >>>> } >>>> >>>> IMHO, maybe candidate devices should have different priorities? For example, the loopback device >>>> may be preferred to use if loopback is available. >>>> >>> IMO, I'd prefer such a order: ISM -> loopback -> RoCE >>> Because ISM for SMC-D is our standard user case, not loopback. >> >> OK, will follow this order. > > My initial thought would be: loopback -> ISM -> RoCE, > just as it is for netdev loopback. > But that only makes sense if loopback performs better than ISM. > Can we postpone that decision until we have measurements? > Sure, it's very reasonable.
Hi, all # Background As previously mentioned in [1], we (Alibaba Cloud) are trying to use SMC to accelerate TCP applications in cloud environment, improving inter-host or inter-VM communication. In addition of these, we also found the value of SMC-D in scenario of local inter-process communication, such as accelerate communication between containers within the same host. So this RFC tries to provide a SMC-D loopback solution in such scenario, to bring a significant improvement in latency and throughput compared to TCP loopback. # Design This patch set provides a kind of SMC-D loopback solution. Patch #1/5 and #2/5 provide an SMC-D based dummy device, preparing for the inter-process communication acceleration. Except for loopback acceleration, the dummy device can also meet the requirements mentioned in [2], which is providing a way to test SMC-D logic for broad community without ISM device. +------------------------------------------+ | +-----------+ +-----------+ | | | process A | | process B | | | +-----------+ +-----------+ | | ^ ^ | | | +---------------+ | | | | | SMC stack | | | | +--->| +-----------+ |<--| | | | | dummy | | | | | | device | | | | +-+-----------+-+ | | VM | +------------------------------------------+ Patch #3/5, #4/5, #5/5 provides a way to avoid data copy from sndbuf to RMB and improve SMC-D loopback performance. Through extending smcd_ops with two new semantic: attach_dmb and detach_dmb, sender's sndbuf shares the same physical memory region with receiver's RMB. The data copied from userspace to sender's sndbuf directly reaches the receiver's RMB without unnecessary memory copy in the same kernel. +----------+ +----------+ | socket A | | socket B | +----------+ +----------+ | ^ | +---------+ | regard as | | ----------| local sndbuf | B's | regard as | | RMB | local RMB |-------> | | +---------+ # Benchmark Test * Test environments: - VM with Intel Xeon Platinum 8 core 2.50GHz, 16 GiB mem. - SMC sndbuf/RMB size 1MB. * Test object: - TCP: run on TCP loopback. - domain: run on UNIX domain. - SMC lo: run on SMC loopback device with patch #1/5 ~ #2/5. - SMC lo-nocpy: run on SMC loopback device with patch #1/5 ~ #5/5. 1. ipc-benchmark (see [3]) - ./<foo> -c 1000000 -s 100 TCP domain SMC-lo SMC-lo-nocpy Message rate (msg/s) 75140 129548(+72.41) 152266(+102.64%) 151914(+102.17%) 2. sockperf - serv: <smc_run> taskset -c <cpu> sockperf sr --tcp - clnt: <smc_run> taskset -c <cpu> sockperf { tp | pp } --tcp --msg-size={ 64000 for tp | 14 for pp } -i 127.0.0.1 -t 30 TCP SMC-lo SMC-lo-nocpy Bandwidth(MBps) 4943.359 4936.096(-0.15%) 8239.624(+66.68%) Latency(us) 6.372 3.359(-47.28%) 3.25(-49.00%) 3. iperf3 - serv: <smc_run> taskset -c <cpu> iperf3 -s - clnt: <smc_run> taskset -c <cpu> iperf3 -c 127.0.0.1 -t 15 TCP SMC-lo SMC-lo-nocpy Bitrate(Gb/s) 40.5 41.4(+2.22%) 76.4(+88.64%) 4. nginx/wrk - serv: <smc_run> nginx - clnt: <smc_run> wrk -t 8 -c 500 -d 30 http://127.0.0.1:80 TCP SMC-lo SMC-lo-nocpy Requests/s 154643.22 220894.03(+42.84%) 226754.3(+46.63%) # Discussion 1. API between SMC-D and ISM device As Jan mentioned in [2], IBM are working on placing an API between SMC-D and the ISM device for easier use of different "devices" for SMC-D. So, considering that the introduction of attach_dmb or detach_dmb can effectively avoid data copying from sndbuf to RMB and brings obvious throughput advantages in inter-VM or inter-process scenarios, can the attach/detach semantics be taken into consideration when designing the API to make it a standard ISM device behavior? Maybe our RFC of SMC-D based inter-process acceleration (this one) and inter-VM acceleration (will coming soon, which is the update of [1]) can provide some examples for new API design. And we are very glad to discuss this on the mail list. 2. Way to select different ISM-like devices With the proposal of SMC-D loopback 'device' (this RFC) and incoming device used for inter-VM acceleration as update of [1], SMC-D has more options to choose from. So we need to consider that how to indicate supported devices, how to determine which one to use, and their priority... IMHO, this may require an update of CLC message and negotiation mechanism. Again, we are very glad to discuss this with you on the mailing list. [1] https://lore.kernel.org/netdev/20220720170048.20806-1-tonylu@linux.alibaba.com/ [2] https://lore.kernel.org/netdev/35d14144-28f7-6129-d6d3-ba16dae7a646@linux.ibm.com/ [3] https://github.com/goldsborough/ipc-bench v1->v2 1. Fix some build WARNINGs complained by kernel test rebot Reported-by: kernel test robot <lkp@intel.com> 2. Add iperf3 test data. Wen Gu (5): net/smc: introduce SMC-D loopback device net/smc: choose loopback device in SMC-D communication net/smc: add dmb attach and detach interface net/smc: avoid data copy from sndbuf to peer RMB in SMC-D loopback net/smc: logic of cursors update in SMC-D loopback connections include/net/smc.h | 3 + net/smc/Makefile | 2 +- net/smc/af_smc.c | 88 +++++++++++- net/smc/smc_cdc.c | 59 ++++++-- net/smc/smc_cdc.h | 1 + net/smc/smc_clc.c | 4 +- net/smc/smc_core.c | 62 +++++++++ net/smc/smc_core.h | 2 + net/smc/smc_ism.c | 39 +++++- net/smc/smc_ism.h | 2 + net/smc/smc_loopback.c | 358 +++++++++++++++++++++++++++++++++++++++++++++++++ net/smc/smc_loopback.h | 63 +++++++++ 12 files changed, 662 insertions(+), 21 deletions(-) create mode 100644 net/smc/smc_loopback.c create mode 100644 net/smc/smc_loopback.h