| Message ID | 20181105165558.11698-1-daniel.m.jordan@oracle.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | ktask: multithread CPU-intensive kernel work | expand |
On Mon 05-11-18 11:55:45, Daniel Jordan wrote: > Michal, you mentioned that ktask should be sensitive to CPU utilization[1]. > ktask threads now run at the lowest priority on the system to avoid disturbing > busy CPUs (more details in patches 4 and 5). Does this address your concern? > The plan to address your other comments is explained below. I have only glanced through the documentation patch and it looks like it will be much less disruptive than the previous attempts. Now the obvious question is how does this behave on a moderately or even busy system when you compare that to a single threaded execution. Some numbers about best/worst case execution would be really helpful. I will look closer later.
Hi Daniel, On 5 Nov 2018, at 11:55, Daniel Jordan wrote: > Hi, > > This version addresses some of the feedback from Andrew and Michal last year > and describes the plan for tackling the rest. I'm posting now since I'll be > presenting ktask at Plumbers next week. > > Andrew, you asked about parallelizing in more places[0]. This version adds > multithreading for VFIO page pinning, and there are more planned users listed > below. > > Michal, you mentioned that ktask should be sensitive to CPU utilization[1]. > ktask threads now run at the lowest priority on the system to avoid disturbing > busy CPUs (more details in patches 4 and 5). Does this address your concern? > The plan to address your other comments is explained below. > > Alex, any thoughts about the VFIO changes in patches 6-9? > > Tejun and Lai, what do you think of patch 5? > > And for everyone, questions and comments welcome. Any suggestions for more > users? > > Thanks, > Daniel > > P.S. This series is big to address the above feedback, but I can send patches > 7 and 8 separately. > > > TODO > ---- > > - Implement cgroup-aware unbound workqueues in a separate series, picking up > Bandan Das's effort from two years ago[2]. This should hopefully address > Michal's comment about running ktask threads within the limits of the calling > context[1]. > > - Make ktask aware of power management. A starting point is to disable the > framework when energy-conscious cpufreq settings are enabled (e.g. > powersave, conservative scaling governors). This should address another > comment from Michal about keeping CPUs under power constraints idle[1]. > > - Add more users. On my list: > - __ib_umem_release in IB core, which Jason Gunthorpe mentioned[3] > - XFS quotacheck and online repair, as suggested by Darrick Wong > - vfs object teardown at umount time, as Andrew mentioned[0] > - page freeing in munmap/exit, as Aaron Lu posted[4] > - page freeing in shmem > The last three will benefit from scaling zone->lock and lru_lock. > > - CPU hotplug support for ktask to adjust its per-CPU data and resource > limits. > > - Check with IOMMU folks that iommu_map is safe for all IOMMU backend > implementations (it is for x86). > > > Summary > ------- > > A single CPU can spend an excessive amount of time in the kernel operating > on large amounts of data. Often these situations arise during initialization- > and destruction-related tasks, where the data involved scales with system size. > These long-running jobs can slow startup and shutdown of applications and the > system itself while extra CPUs sit idle. > > To ensure that applications and the kernel continue to perform well as core > counts and memory sizes increase, harness these idle CPUs to complete such jobs > more quickly. > > ktask is a generic framework for parallelizing CPU-intensive work in the > kernel. The API is generic enough to add concurrency to many different kinds > of tasks--for example, zeroing a range of pages or evicting a list of > inodes--and aims to save its clients the trouble of splitting up the work, > choosing the number of threads to use, maintaining an efficient concurrency > level, starting these threads, and load balancing the work between them. > > The first patch has more documentation, and the second patch has the interface. > > Current users: > 1) VFIO page pinning before kvm guest startup (others hitting slowness too[5]) > 2) deferred struct page initialization at boot time > 3) clearing gigantic pages > 4) fallocate for HugeTLB pages Do you think if it makes sense to use ktask for huge page migration (the data copy part)? I did some experiments back in 2016[1], which showed that migrating one 2MB page with 8 threads could achieve 2.8x throughput of the existing single-threaded method. The problem with my parallel page migration patchset at that time was that it has no CPU-utilization awareness, which is solved by your patches now. Thanks. [1]https://lkml.org/lkml/2016/11/22/457 -- Best Regards Yan Zi
On Mon, Nov 05, 2018 at 06:29:31PM +0100, Michal Hocko wrote: > On Mon 05-11-18 11:55:45, Daniel Jordan wrote: > > Michal, you mentioned that ktask should be sensitive to CPU utilization[1]. > > ktask threads now run at the lowest priority on the system to avoid disturbing > > busy CPUs (more details in patches 4 and 5). Does this address your concern? > > The plan to address your other comments is explained below. > > I have only glanced through the documentation patch and it looks like it > will be much less disruptive than the previous attempts. Now the obvious > question is how does this behave on a moderately or even busy system > when you compare that to a single threaded execution. Some numbers about > best/worst case execution would be really helpful. Patches 4 and 5 have some numbers where a ktask and non-ktask workload compete against each other. Those show either 8 ktask threads on 8 CPUs (worst case) or no ktask threads (best case). By single threaded execution, I guess you mean 1 ktask thread. I'll run the experiments that way too and post the numbers. > I will look closer later. Great! Thanks for your comment. Daniel
Hi Zi, On Mon, Nov 05, 2018 at 01:49:14PM -0500, Zi Yan wrote: > On 5 Nov 2018, at 11:55, Daniel Jordan wrote: > > Do you think if it makes sense to use ktask for huge page migration (the data > copy part)? It certainly could. > I did some experiments back in 2016[1], which showed that migrating one 2MB page > with 8 threads could achieve 2.8x throughput of the existing single-threaded method. > The problem with my parallel page migration patchset at that time was that it > has no CPU-utilization awareness, which is solved by your patches now. Did you run with fewer than 8 threads? I'd want a bigger speedup than 2.8x for 8, and a smaller thread count might improve thread utilization. It would be nice to multithread at a higher granularity than 2M, too: a range of THPs might also perform better than a single page. Thanks for your comments. > [1]https://lkml.org/lkml/2016/11/22/457
On 5 Nov 2018, at 21:20, Daniel Jordan wrote: > Hi Zi, > > On Mon, Nov 05, 2018 at 01:49:14PM -0500, Zi Yan wrote: >> On 5 Nov 2018, at 11:55, Daniel Jordan wrote: >> >> Do you think if it makes sense to use ktask for huge page migration (the data >> copy part)? > > It certainly could. > >> I did some experiments back in 2016[1], which showed that migrating one 2MB page >> with 8 threads could achieve 2.8x throughput of the existing single-threaded method. >> The problem with my parallel page migration patchset at that time was that it >> has no CPU-utilization awareness, which is solved by your patches now. > > Did you run with fewer than 8 threads? I'd want a bigger speedup than 2.8x for > 8, and a smaller thread count might improve thread utilization. Yes. When migrating one 2MB THP with migrate_pages() system call on a two-socket server with 2 E5-2650 v3 CPUs (10 cores per socket) across two sockets, here are the page migration throughput numbers: throughput factor 1 thread 2.15 GB/s 1x 2 threads 3.05 GB/s 1.42x 4 threads 4.50 GB/s 2.09x 8 threads 5.98 GB/s 2.78x > > It would be nice to multithread at a higher granularity than 2M, too: a range > of THPs might also perform better than a single page. Sure. But the kernel currently does not copy multiple pages altogether even if a range of THPs is migrated. Page copy function is interleaved with page table operations for every single page. I also did some study and modified the kernel to improve this, which I called concurrent page migration in https://lwn.net/Articles/714991/. It further improves page migration throughput. — Best Regards, Yan Zi
On Mon 05-11-18 17:29:55, Daniel Jordan wrote: > On Mon, Nov 05, 2018 at 06:29:31PM +0100, Michal Hocko wrote: > > On Mon 05-11-18 11:55:45, Daniel Jordan wrote: > > > Michal, you mentioned that ktask should be sensitive to CPU utilization[1]. > > > ktask threads now run at the lowest priority on the system to avoid disturbing > > > busy CPUs (more details in patches 4 and 5). Does this address your concern? > > > The plan to address your other comments is explained below. > > > > I have only glanced through the documentation patch and it looks like it > > will be much less disruptive than the previous attempts. Now the obvious > > question is how does this behave on a moderately or even busy system > > when you compare that to a single threaded execution. Some numbers about > > best/worst case execution would be really helpful. > > Patches 4 and 5 have some numbers where a ktask and non-ktask workload compete > against each other. Those show either 8 ktask threads on 8 CPUs (worst case) or no ktask threads (best case). > > By single threaded execution, I guess you mean 1 ktask thread. I'll run the > experiments that way too and post the numbers. I mean a comparision of how much time it gets to accomplish the same amount of work if it was done singlethreaded to ktask based distribution on a idle system (best case for both) and fully contended system (the worst case). It would be also great to get some numbers on partially contended system to see how much the priority handover etc. acts under different CPU contention.
On Mon, Nov 05, 2018 at 09:48:56PM -0500, Zi Yan wrote: > On 5 Nov 2018, at 21:20, Daniel Jordan wrote: > > > Hi Zi, > > > > On Mon, Nov 05, 2018 at 01:49:14PM -0500, Zi Yan wrote: > >> On 5 Nov 2018, at 11:55, Daniel Jordan wrote: > >> > >> Do you think if it makes sense to use ktask for huge page migration (the data > >> copy part)? > > > > It certainly could. > > > >> I did some experiments back in 2016[1], which showed that migrating one 2MB page > >> with 8 threads could achieve 2.8x throughput of the existing single-threaded method. > >> The problem with my parallel page migration patchset at that time was that it > >> has no CPU-utilization awareness, which is solved by your patches now. > > > > Did you run with fewer than 8 threads? I'd want a bigger speedup than 2.8x for > > 8, and a smaller thread count might improve thread utilization. > > Yes. When migrating one 2MB THP with migrate_pages() system call on a two-socket server > with 2 E5-2650 v3 CPUs (10 cores per socket) across two sockets, here are the page migration > throughput numbers: > > throughput factor > 1 thread 2.15 GB/s 1x > 2 threads 3.05 GB/s 1.42x > 4 threads 4.50 GB/s 2.09x > 8 threads 5.98 GB/s 2.78x Thanks. Looks like in your patches you start a worker for every piece of the huge page copy and have the main thread wait. I'm curious what the workqueue overhead is like on your machine. On a newer Xeon it's ~50usec from queueing a work to starting to execute it and another ~20usec to flush a work (barrier_func), which could happen after the work is already done. A pretty significant piece of the copy time for part of a THP. bash 60728 [087] 155865.157116: probe:ktask_run: (ffffffffb7ee7a80) bash 60728 [087] 155865.157119: workqueue:workqueue_queue_work: work struct=0xffff95fb73276000 bash 60728 [087] 155865.157119: workqueue:workqueue_activate_work: work struct 0xffff95fb73276000 kworker/u194:3- 86730 [095] 155865.157168: workqueue:workqueue_execute_start: work struct 0xffff95fb73276000: function ktask_thread kworker/u194:3- 86730 [095] 155865.157170: workqueue:workqueue_execute_end: work struct 0xffff95fb73276000 kworker/u194:3- 86730 [095] 155865.157171: workqueue:workqueue_execute_start: work struct 0xffffa676995bfb90: function wq_barrier_func kworker/u194:3- 86730 [095] 155865.157190: workqueue:workqueue_execute_end: work struct 0xffffa676995bfb90 bash 60728 [087] 155865.157207: probe:ktask_run_ret__return: (ffffffffb7ee7a80 <- ffffffffb7ee7b7b) > > > > It would be nice to multithread at a higher granularity than 2M, too: a range > > of THPs might also perform better than a single page. > > Sure. But the kernel currently does not copy multiple pages altogether even if a range > of THPs is migrated. Page copy function is interleaved with page table operations > for every single page. > > I also did some study and modified the kernel to improve this, which I called > concurrent page migration in https://lwn.net/Articles/714991/. It further > improves page migration throughput. Ok, over 4x with 8 threads for 16 THPs. Is 16 a typical number for migration, or does it get larger? What workloads do you have in mind with this change?
On Tue, Nov 06, 2018 at 10:21:45AM +0100, Michal Hocko wrote: > On Mon 05-11-18 17:29:55, Daniel Jordan wrote: > > On Mon, Nov 05, 2018 at 06:29:31PM +0100, Michal Hocko wrote: > > > On Mon 05-11-18 11:55:45, Daniel Jordan wrote: > > > > Michal, you mentioned that ktask should be sensitive to CPU utilization[1]. > > > > ktask threads now run at the lowest priority on the system to avoid disturbing > > > > busy CPUs (more details in patches 4 and 5). Does this address your concern? > > > > The plan to address your other comments is explained below. > > > > > > I have only glanced through the documentation patch and it looks like it > > > will be much less disruptive than the previous attempts. Now the obvious > > > question is how does this behave on a moderately or even busy system > > > when you compare that to a single threaded execution. Some numbers about > > > best/worst case execution would be really helpful. > > > > Patches 4 and 5 have some numbers where a ktask and non-ktask workload compete > > against each other. Those show either 8 ktask threads on 8 CPUs (worst case) or no ktask threads (best case). > > > > By single threaded execution, I guess you mean 1 ktask thread. I'll run the > > experiments that way too and post the numbers. > > I mean a comparision of how much time it gets to accomplish the same > amount of work if it was done singlethreaded to ktask based distribution > on a idle system (best case for both) and fully contended system (the > worst case). It would be also great to get some numbers on partially > contended system to see how much the priority handover etc. acts under > different CPU contention. Ok, thanks for clarifying. Testing notes - The two workloads used were confined to run anywhere within an 8-CPU cpumask - The vfio workload started a 64G VM using THP - usemem was enlisted to create CPU load doing page clearing, just as the vfio case is doing, so the two compete for the same system resources. usemem ran four times with each of its threads allocating and freeing 30G of memory each time. Four usemem threads simulate Michal's partially contended system - ktask helpers always run at MAX_NICE - renice?=yes means run with patch 5, renice?=no means without - CPU: 2 nodes * 24 cores/node * 2 threads/core = 96 CPUs Intel(R) Xeon(R) Platinum 8160 CPU @ 2.10GHz vfio usemem thr thr renice? ktask sec usemem sec ----- ------ ------- ---------------- ---------------- 4 n/a 24.0 ( ± 0.1% ) 8 n/a 25.3 ( ± 0.0% ) 1 0 n/a 13.5 ( ± 0.0% ) 1 4 n/a 14.2 ( ± 0.4% ) 24.1 ( ± 0.3% ) *** 1 8 n/a 17.3 ( ± 10.4% ) 29.7 ( ± 0.4% ) 8 0 no 2.8 ( ± 1.5% ) 8 4 no 4.7 ( ± 0.8% ) 24.1 ( ± 0.2% ) 8 8 no 13.7 ( ± 8.8% ) 27.2 ( ± 1.2% ) 8 0 yes 2.8 ( ± 1.0% ) 8 4 yes 4.7 ( ± 1.4% ) 24.1 ( ± 0.0% ) *** 8 8 yes 9.2 ( ± 2.2% ) 27.0 ( ± 0.4% ) Renicing under partial contention (usemem nthr=4) doesn't affect vfio, but renicing under heavy contention (usemem nthr=8) does: the 8-thread vfio case is slower when the ktask master thread doesn't will its priority to each helper at a time. Comparing the ***'d lines, using 8 vfio threads instead of 1 causes the threads of both workloads to finish sooner under heavy contention.
Hello, On Mon, Nov 05, 2018 at 11:55:45AM -0500, Daniel Jordan wrote: > Michal, you mentioned that ktask should be sensitive to CPU utilization[1]. > ktask threads now run at the lowest priority on the system to avoid disturbing > busy CPUs (more details in patches 4 and 5). Does this address your concern? > The plan to address your other comments is explained below. Have you tested what kind of impact this has on bandwidth of a system in addition to latency? The thing is while this would make a better use of a system which has idle capacity, it does so by doing more total work. It'd be really interesting to see how this affects bandwidth of a system too. Thanks.
On Fri, Nov 30, 2018 at 11:18:19AM -0800, Tejun Heo wrote: > Hello, > > On Mon, Nov 05, 2018 at 11:55:45AM -0500, Daniel Jordan wrote: > > Michal, you mentioned that ktask should be sensitive to CPU utilization[1]. > > ktask threads now run at the lowest priority on the system to avoid disturbing > > busy CPUs (more details in patches 4 and 5). Does this address your concern? > > The plan to address your other comments is explained below. > > Have you tested what kind of impact this has on bandwidth of a system > in addition to latency? The thing is while this would make a better > use of a system which has idle capacity, it does so by doing more > total work. It'd be really interesting to see how this affects > bandwidth of a system too. I guess you mean something like comparing aggregate CPU time across threads to the base single thread time for some job or set of jobs? Then no, I haven't measured that, but I can for next time.
Hello, On Fri, Nov 30, 2018 at 04:13:07PM -0800, Daniel Jordan wrote: > On Fri, Nov 30, 2018 at 11:18:19AM -0800, Tejun Heo wrote: > > Hello, > > > > On Mon, Nov 05, 2018 at 11:55:45AM -0500, Daniel Jordan wrote: > > > Michal, you mentioned that ktask should be sensitive to CPU utilization[1]. > > > ktask threads now run at the lowest priority on the system to avoid disturbing > > > busy CPUs (more details in patches 4 and 5). Does this address your concern? > > > The plan to address your other comments is explained below. > > > > Have you tested what kind of impact this has on bandwidth of a system > > in addition to latency? The thing is while this would make a better > > use of a system which has idle capacity, it does so by doing more > > total work. It'd be really interesting to see how this affects > > bandwidth of a system too. > > I guess you mean something like comparing aggregate CPU time across threads to > the base single thread time for some job or set of jobs? Then no, I haven't > measured that, but I can for next time. Yeah, I'm primarily curious how expensive this is on an already loaded system, so sth like loading up the system with a workload which can saturate the system and comparing the bw impacts of serial and parallel page clearings at the same frequency. Thanks.
Hi, This version addresses some of the feedback from Andrew and Michal last year and describes the plan for tackling the rest. I'm posting now since I'll be presenting ktask at Plumbers next week. Andrew, you asked about parallelizing in more places[0]. This version adds multithreading for VFIO page pinning, and there are more planned users listed below. Michal, you mentioned that ktask should be sensitive to CPU utilization[1]. ktask threads now run at the lowest priority on the system to avoid disturbing busy CPUs (more details in patches 4 and 5). Does this address your concern? The plan to address your other comments is explained below. Alex, any thoughts about the VFIO changes in patches 6-9? Tejun and Lai, what do you think of patch 5? And for everyone, questions and comments welcome. Any suggestions for more users? Thanks, Daniel P.S. This series is big to address the above feedback, but I can send patches 7 and 8 separately. TODO ---- - Implement cgroup-aware unbound workqueues in a separate series, picking up Bandan Das's effort from two years ago[2]. This should hopefully address Michal's comment about running ktask threads within the limits of the calling context[1]. - Make ktask aware of power management. A starting point is to disable the framework when energy-conscious cpufreq settings are enabled (e.g. powersave, conservative scaling governors). This should address another comment from Michal about keeping CPUs under power constraints idle[1]. - Add more users. On my list: - __ib_umem_release in IB core, which Jason Gunthorpe mentioned[3] - XFS quotacheck and online repair, as suggested by Darrick Wong - vfs object teardown at umount time, as Andrew mentioned[0] - page freeing in munmap/exit, as Aaron Lu posted[4] - page freeing in shmem The last three will benefit from scaling zone->lock and lru_lock. - CPU hotplug support for ktask to adjust its per-CPU data and resource limits. - Check with IOMMU folks that iommu_map is safe for all IOMMU backend implementations (it is for x86). Summary ------- A single CPU can spend an excessive amount of time in the kernel operating on large amounts of data. Often these situations arise during initialization- and destruction-related tasks, where the data involved scales with system size. These long-running jobs can slow startup and shutdown of applications and the system itself while extra CPUs sit idle. To ensure that applications and the kernel continue to perform well as core counts and memory sizes increase, harness these idle CPUs to complete such jobs more quickly. ktask is a generic framework for parallelizing CPU-intensive work in the kernel. The API is generic enough to add concurrency to many different kinds of tasks--for example, zeroing a range of pages or evicting a list of inodes--and aims to save its clients the trouble of splitting up the work, choosing the number of threads to use, maintaining an efficient concurrency level, starting these threads, and load balancing the work between them. The first patch has more documentation, and the second patch has the interface. Current users: 1) VFIO page pinning before kvm guest startup (others hitting slowness too[5]) 2) deferred struct page initialization at boot time 3) clearing gigantic pages 4) fallocate for HugeTLB pages This patchset is based on the 2018-10-30 head of mmotm/master. Changelog: v3 -> v4: - Added VFIO page pinning use case (Andrew's "more users" comment) - Made ktask helpers run at the lowest priority on the system (Michal's concern about sensitivity to CPU utilization) - ktask learned to undo part of a task on error (required for VFIO) - Changed mm->locked_vm to an atomic to improve performance for VFIO. This can be split out into a smaller series (there wasn't time before posting this) - Removed /proc/sys/debug/ktask_max_threads (it was a debug-only thing) - Some minor improvements in the ktask code itself (shorter, cleaner, etc) - Updated Documentation to cover these changes v2 -> v3: - Changed cpu to CPU in the ktask Documentation, as suggested by Randy Dunlap - Saved more boot time now that Pavel Tatashin's deferred struct page init patches are in mainline (https://lkml.org/lkml/2017/10/13/692). New performance results in patch 7. - Added resource limits, per-node and system-wide, to maintain efficient concurrency levels (addresses a concern from my Plumbers talk) - ktask no longer allocates memory internally during a task so it can be used in sensitive contexts - Added the option to run work anywhere on the system rather than always confining it to a specific node - Updated Documentation patch with these changes and reworked motivation section v1 -> v2: - Added deferred struct page initialization use case. - Explained the source of the performance improvement from parallelizing clear_gigantic_page (comment from Dave Hansen). - Fixed Documentation and build warnings from CONFIG_KTASK=n kernels. ktask v3 RFC: lkml.kernel.org/r/20171205195220.28208-1-daniel.m.jordan@oracle.com [0] https://lkml.kernel.org/r/20171205142300.67489b1a90605e1089c5aaa9@linux-foundation.org [1] https://lkml.kernel.org/r/20171206143509.GG7515@dhcp22.suse.cz [2] https://lkml.kernel.org/r/1458339291-4093-1-git-send-email-bsd@redhat.com [3] https://lkml.kernel.org/r/20180928153922.GA17076@ziepe.ca [4] https://lkml.kernel.org/r/1489568404-7817-1-git-send-email-aaron.lu@intel.com [5] https://www.redhat.com/archives/vfio-users/2018-April/msg00020.html Daniel Jordan (13): ktask: add documentation ktask: multithread CPU-intensive kernel work ktask: add undo support ktask: run helper threads at MAX_NICE workqueue, ktask: renice helper threads to prevent starvation vfio: parallelize vfio_pin_map_dma mm: change locked_vm's type from unsigned long to atomic_long_t vfio: remove unnecessary mmap_sem writer acquisition around locked_vm vfio: relieve mmap_sem reader cacheline bouncing by holding it longer mm: enlarge type of offset argument in mem_map_offset and mem_map_next mm: parallelize deferred struct page initialization within each node mm: parallelize clear_gigantic_page hugetlbfs: parallelize hugetlbfs_fallocate with ktask Documentation/core-api/index.rst | 1 + Documentation/core-api/ktask.rst | 213 +++++++++ arch/powerpc/kvm/book3s_64_vio.c | 15 +- arch/powerpc/mm/mmu_context_iommu.c | 16 +- drivers/fpga/dfl-afu-dma-region.c | 16 +- drivers/vfio/vfio_iommu_spapr_tce.c | 14 +- drivers/vfio/vfio_iommu_type1.c | 159 ++++--- fs/hugetlbfs/inode.c | 114 ++++- fs/proc/task_mmu.c | 2 +- include/linux/ktask.h | 267 ++++++++++++ include/linux/mm_types.h | 2 +- include/linux/workqueue.h | 5 + init/Kconfig | 11 + init/main.c | 2 + kernel/Makefile | 2 +- kernel/fork.c | 2 +- kernel/ktask.c | 646 ++++++++++++++++++++++++++++ kernel/workqueue.c | 106 ++++- mm/debug.c | 3 +- mm/internal.h | 7 +- mm/memory.c | 32 +- mm/mlock.c | 4 +- mm/mmap.c | 18 +- mm/mremap.c | 6 +- mm/page_alloc.c | 91 +++- 25 files changed, 1599 insertions(+), 155 deletions(-) create mode 100644 Documentation/core-api/ktask.rst create mode 100644 include/linux/ktask.h create mode 100644 kernel/ktask.c