| Message ID | 20240709160551.40595-1-guangguan.wang@linux.alibaba.com (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [net-next] net/smc: introduce autosplit for smc | expand |
On 09.07.24 18:05, Guangguan Wang wrote: > When sending large size data in TCP, the data will be split into > several segments(packets) to transfer due to MTU config. And in > the receive side, application can be woken up to recv data every > packet arrived, the data transmission and data recv copy are > pipelined. > > But for SMC-R, it will transmit as many data as possible in one > RDMA WRITE and a CDC msg follows the RDMA WRITE, in the receive > size, the application only be woken up to recv data when all RDMA > WRITE data and the followed CDC msg arrived. The data transmission > and data recv copy are sequential. > > This patch introduce autosplit for SMC, which can automatic split > data into several segments and every segment transmitted by one RDMA > WRITE when sending large size data in SMC. Because of the split, the > data transmission and data send copy can be pipelined in the send side, > and the data transmission and data recv copy can be pipelined in the > receive side. Thus autosplit helps improving latency performance when > sending large size data. The autosplit also works for SMC-D. > > This patch also introduce a sysctl names autosplit_size for configure > the max size of the split segment, whose default value is 128KiB > (128KiB perform best in my environment). > > The sockperf benchmark shows 17%-28% latency improvement when msgsize >> = 256KB for SMC-R, 15%-32% latency improvement when msgsize >= 256KB > for SMC-D with smc-loopback. > > Test command: > sockperf sr --tcp -m 1048575 > sockperf pp --tcp -i <server ip> -m <msgsize> -t 20 > > Test config: > sysctl -w net.smc.wmem=524288 > sysctl -w net.smc.rmem=524288 > > Test results: > SMC-R > msgsize noautosplit autosplit > 128KB 55.546 us 55.763 us > 256KB 83.537 us 69.743 us (17% improve) > 512KB 138.306 us 100.313 us (28% improve) > 1MB 273.702 us 197.222 us (28% improve) > > SMC-D with smc-loopback > msgsize noautosplit autosplit > 128KB 14.672 us 14.690 us > 256KB 28.277 us 23.958 us (15% improve) > 512KB 63.047 us 45.339 us (28% improve) > 1MB 129.306 us 87.278 us (32% improve) > > Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com> > --- > Documentation/networking/smc-sysctl.rst | 11 +++++++++++ > include/net/netns/smc.h | 1 + > net/smc/smc_sysctl.c | 12 ++++++++++++ > net/smc/smc_tx.c | 19 ++++++++++++++++++- > 4 files changed, 42 insertions(+), 1 deletion(-) > Hi Guangguan, If I remember correctly, the intention to use one RDMA-write for a possible large data is to reduce possible many partial stores. Since many year has gone, I'm not that sure if it would still be an issue. I need some time to check on it. BTW, I don't really like the idea to use sysctl to set the autosplit_size in any value at will. That makes no sense to improve the performance. Thanks, Wenjia
On 2024/7/11 23:57, Wenjia Zhang wrote: > > > On 09.07.24 18:05, Guangguan Wang wrote: >> When sending large size data in TCP, the data will be split into >> several segments(packets) to transfer due to MTU config. And in >> the receive side, application can be woken up to recv data every >> packet arrived, the data transmission and data recv copy are >> pipelined. >> >> But for SMC-R, it will transmit as many data as possible in one >> RDMA WRITE and a CDC msg follows the RDMA WRITE, in the receive >> size, the application only be woken up to recv data when all RDMA >> WRITE data and the followed CDC msg arrived. The data transmission >> and data recv copy are sequential. >> >> This patch introduce autosplit for SMC, which can automatic split >> data into several segments and every segment transmitted by one RDMA >> WRITE when sending large size data in SMC. Because of the split, the >> data transmission and data send copy can be pipelined in the send side, >> and the data transmission and data recv copy can be pipelined in the >> receive side. Thus autosplit helps improving latency performance when >> sending large size data. The autosplit also works for SMC-D. >> >> This patch also introduce a sysctl names autosplit_size for configure >> the max size of the split segment, whose default value is 128KiB >> (128KiB perform best in my environment). >> >> The sockperf benchmark shows 17%-28% latency improvement when msgsize >>> = 256KB for SMC-R, 15%-32% latency improvement when msgsize >= 256KB >> for SMC-D with smc-loopback. >> >> Test command: >> sockperf sr --tcp -m 1048575 >> sockperf pp --tcp -i <server ip> -m <msgsize> -t 20 >> >> Test config: >> sysctl -w net.smc.wmem=524288 >> sysctl -w net.smc.rmem=524288 >> >> Test results: >> SMC-R >> msgsize noautosplit autosplit >> 128KB 55.546 us 55.763 us >> 256KB 83.537 us 69.743 us (17% improve) >> 512KB 138.306 us 100.313 us (28% improve) >> 1MB 273.702 us 197.222 us (28% improve) >> >> SMC-D with smc-loopback >> msgsize noautosplit autosplit >> 128KB 14.672 us 14.690 us >> 256KB 28.277 us 23.958 us (15% improve) >> 512KB 63.047 us 45.339 us (28% improve) >> 1MB 129.306 us 87.278 us (32% improve) >> >> Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com> >> --- >> Documentation/networking/smc-sysctl.rst | 11 +++++++++++ >> include/net/netns/smc.h | 1 + >> net/smc/smc_sysctl.c | 12 ++++++++++++ >> net/smc/smc_tx.c | 19 ++++++++++++++++++- >> 4 files changed, 42 insertions(+), 1 deletion(-) >> > > Hi Guangguan, > > If I remember correctly, the intention to use one RDMA-write for a possible large data is to reduce possible many partial stores. Since many year has gone, I'm not that sure if it would still be an issue. I need some time to check on it. > Did you mean too many partial stores will result in some issue? What's the issue? > BTW, I don't really like the idea to use sysctl to set the autosplit_size in any value at will. That makes no sense to improve the performance. Although 128KB autosplit_size have a good performance in most scenario, I still found some better autosplit_size for some specific network configurations. For example, 128KB autosplit_size have a good performance whether the MTU is 1500 or 8500, but for 8500 MTU, 64KB autosplit_size performs better. Maybe the sysctl is not the best way, but I think it should have a way to set the value of autosplit_size for possible performance tuning. Thanks, Guangguan Wang > > Thanks, > Wenjia
On 2024/7/15 10:53, Guangguan Wang wrote: > > > On 2024/7/11 23:57, Wenjia Zhang wrote: >> >> >> On 09.07.24 18:05, Guangguan Wang wrote: >>> When sending large size data in TCP, the data will be split into >>> several segments(packets) to transfer due to MTU config. And in >>> the receive side, application can be woken up to recv data every >>> packet arrived, the data transmission and data recv copy are >>> pipelined. >>> >>> But for SMC-R, it will transmit as many data as possible in one >>> RDMA WRITE and a CDC msg follows the RDMA WRITE, in the receive >>> size, the application only be woken up to recv data when all RDMA >>> WRITE data and the followed CDC msg arrived. The data transmission >>> and data recv copy are sequential. >>> >>> This patch introduce autosplit for SMC, which can automatic split >>> data into several segments and every segment transmitted by one RDMA >>> WRITE when sending large size data in SMC. Because of the split, the >>> data transmission and data send copy can be pipelined in the send side, >>> and the data transmission and data recv copy can be pipelined in the >>> receive side. Thus autosplit helps improving latency performance when >>> sending large size data. The autosplit also works for SMC-D. >>> >>> This patch also introduce a sysctl names autosplit_size for configure >>> the max size of the split segment, whose default value is 128KiB >>> (128KiB perform best in my environment). >>> >>> The sockperf benchmark shows 17%-28% latency improvement when msgsize >>>> = 256KB for SMC-R, 15%-32% latency improvement when msgsize >= 256KB >>> for SMC-D with smc-loopback. >>> >>> Test command: >>> sockperf sr --tcp -m 1048575 >>> sockperf pp --tcp -i <server ip> -m <msgsize> -t 20 >>> >>> Test config: >>> sysctl -w net.smc.wmem=524288 >>> sysctl -w net.smc.rmem=524288 >>> >>> Test results: >>> SMC-R >>> msgsize noautosplit autosplit >>> 128KB 55.546 us 55.763 us >>> 256KB 83.537 us 69.743 us (17% improve) >>> 512KB 138.306 us 100.313 us (28% improve) >>> 1MB 273.702 us 197.222 us (28% improve) >>> >>> SMC-D with smc-loopback >>> msgsize noautosplit autosplit >>> 128KB 14.672 us 14.690 us >>> 256KB 28.277 us 23.958 us (15% improve) >>> 512KB 63.047 us 45.339 us (28% improve) >>> 1MB 129.306 us 87.278 us (32% improve) >>> >>> Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com> >>> --- >>> Documentation/networking/smc-sysctl.rst | 11 +++++++++++ >>> include/net/netns/smc.h | 1 + >>> net/smc/smc_sysctl.c | 12 ++++++++++++ >>> net/smc/smc_tx.c | 19 ++++++++++++++++++- >>> 4 files changed, 42 insertions(+), 1 deletion(-) >>> >> >> Hi Guangguan, >> >> If I remember correctly, the intention to use one RDMA-write for a possible large data is to reduce possible many partial stores. Since many year has gone, I'm not that sure if it would still be an issue. I need some time to check on it. >> > > Did you mean too many partial stores will result in some issue? What's the issue? > > >> BTW, I don't really like the idea to use sysctl to set the autosplit_size in any value at will. That makes no sense to improve the performance. > > Although 128KB autosplit_size have a good performance in most scenario, I still found some better autosplit_size for some specific network configurations. > For example, 128KB autosplit_size have a good performance whether the MTU is 1500 or 8500, but for 8500 MTU, 64KB autosplit_size performs better. > > Maybe the sysctl is not the best way, but I think it should have a way to set the value of autosplit_size for possible performance tuning. > > Thanks, > Guangguan Wang > Hi Wenjia, Is there any update comment or information about this patch? >> >> Thanks, >> Wenjia
On 08.08.24 08:26, Guangguan Wang wrote: > On 2024/7/15 10:53, Guangguan Wang wrote: >> >> >> On 2024/7/11 23:57, Wenjia Zhang wrote: >>> >>> >>> On 09.07.24 18:05, Guangguan Wang wrote: >>>> When sending large size data in TCP, the data will be split into >>>> several segments(packets) to transfer due to MTU config. And in >>>> the receive side, application can be woken up to recv data every >>>> packet arrived, the data transmission and data recv copy are >>>> pipelined. >>>> >>>> But for SMC-R, it will transmit as many data as possible in one >>>> RDMA WRITE and a CDC msg follows the RDMA WRITE, in the receive >>>> size, the application only be woken up to recv data when all RDMA >>>> WRITE data and the followed CDC msg arrived. The data transmission >>>> and data recv copy are sequential. >>>> >>>> This patch introduce autosplit for SMC, which can automatic split >>>> data into several segments and every segment transmitted by one RDMA >>>> WRITE when sending large size data in SMC. Because of the split, the >>>> data transmission and data send copy can be pipelined in the send side, >>>> and the data transmission and data recv copy can be pipelined in the >>>> receive side. Thus autosplit helps improving latency performance when >>>> sending large size data. The autosplit also works for SMC-D. >>>> >>>> This patch also introduce a sysctl names autosplit_size for configure >>>> the max size of the split segment, whose default value is 128KiB >>>> (128KiB perform best in my environment). >>>> >>>> The sockperf benchmark shows 17%-28% latency improvement when msgsize >>>>> = 256KB for SMC-R, 15%-32% latency improvement when msgsize >= 256KB >>>> for SMC-D with smc-loopback. >>>> >>>> Test command: >>>> sockperf sr --tcp -m 1048575 >>>> sockperf pp --tcp -i <server ip> -m <msgsize> -t 20 >>>> >>>> Test config: >>>> sysctl -w net.smc.wmem=524288 >>>> sysctl -w net.smc.rmem=524288 >>>> >>>> Test results: >>>> SMC-R >>>> msgsize noautosplit autosplit >>>> 128KB 55.546 us 55.763 us >>>> 256KB 83.537 us 69.743 us (17% improve) >>>> 512KB 138.306 us 100.313 us (28% improve) >>>> 1MB 273.702 us 197.222 us (28% improve) >>>> >>>> SMC-D with smc-loopback >>>> msgsize noautosplit autosplit >>>> 128KB 14.672 us 14.690 us >>>> 256KB 28.277 us 23.958 us (15% improve) >>>> 512KB 63.047 us 45.339 us (28% improve) >>>> 1MB 129.306 us 87.278 us (32% improve) >>>> >>>> Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com> >>>> --- >>>> Documentation/networking/smc-sysctl.rst | 11 +++++++++++ >>>> include/net/netns/smc.h | 1 + >>>> net/smc/smc_sysctl.c | 12 ++++++++++++ >>>> net/smc/smc_tx.c | 19 ++++++++++++++++++- >>>> 4 files changed, 42 insertions(+), 1 deletion(-) >>>> >>> >>> Hi Guangguan, >>> >>> If I remember correctly, the intention to use one RDMA-write for a possible large data is to reduce possible many partial stores. Since many year has gone, I'm not that sure if it would still be an issue. I need some time to check on it. >>> >> >> Did you mean too many partial stores will result in some issue? What's the issue? >> Forget it, I did verify that the partial stores should not be problem now. >> >>> BTW, I don't really like the idea to use sysctl to set the autosplit_size in any value at will. That makes no sense to improve the performance. >> >> Although 128KB autosplit_size have a good performance in most scenario, I still found some better autosplit_size for some specific network configurations. >> For example, 128KB autosplit_size have a good performance whether the MTU is 1500 or 8500, but for 8500 MTU, 64KB autosplit_size performs better. >> >> Maybe the sysctl is not the best way, but I think it should have a way to set the value of autosplit_size for possible performance tuning. >> mhhh, that could be a good reason to use sysctl. >> Thanks, >> Guangguan Wang >> > > Hi Wenjia, > > Is there any update comment or information about this patch? Hi Guangguan, sorry for the delayed answer. In the last time it is really difficult for me to find time to look into it and test it. With more thinking, I'm kind of convinced with this idea. But test is still needed. I'll be in vacation next 3 weeks. I hope it is okay for you that I'll test it as soon as possible when I'm back. If everything is ok, I think we can let it go upstream. Thanks, Wenjia > >>> >>> Thanks, >>> Wenjia
On 2024/8/10 05:07, Wenjia Zhang wrote: > > > On 08.08.24 08:26, Guangguan Wang wrote: >> On 2024/7/15 10:53, Guangguan Wang wrote: >>> >>> >>> On 2024/7/11 23:57, Wenjia Zhang wrote: >>>> >>>> >>>> On 09.07.24 18:05, Guangguan Wang wrote: >>>>> When sending large size data in TCP, the data will be split into >>>>> several segments(packets) to transfer due to MTU config. And in >>>>> the receive side, application can be woken up to recv data every >>>>> packet arrived, the data transmission and data recv copy are >>>>> pipelined. >>>>> >>>>> But for SMC-R, it will transmit as many data as possible in one >>>>> RDMA WRITE and a CDC msg follows the RDMA WRITE, in the receive >>>>> size, the application only be woken up to recv data when all RDMA >>>>> WRITE data and the followed CDC msg arrived. The data transmission >>>>> and data recv copy are sequential. >>>>> >>>>> This patch introduce autosplit for SMC, which can automatic split >>>>> data into several segments and every segment transmitted by one RDMA >>>>> WRITE when sending large size data in SMC. Because of the split, the >>>>> data transmission and data send copy can be pipelined in the send side, >>>>> and the data transmission and data recv copy can be pipelined in the >>>>> receive side. Thus autosplit helps improving latency performance when >>>>> sending large size data. The autosplit also works for SMC-D. >>>>> >>>>> This patch also introduce a sysctl names autosplit_size for configure >>>>> the max size of the split segment, whose default value is 128KiB >>>>> (128KiB perform best in my environment). >>>>> >>>>> The sockperf benchmark shows 17%-28% latency improvement when msgsize >>>>>> = 256KB for SMC-R, 15%-32% latency improvement when msgsize >= 256KB >>>>> for SMC-D with smc-loopback. >>>>> >>>>> Test command: >>>>> sockperf sr --tcp -m 1048575 >>>>> sockperf pp --tcp -i <server ip> -m <msgsize> -t 20 >>>>> >>>>> Test config: >>>>> sysctl -w net.smc.wmem=524288 >>>>> sysctl -w net.smc.rmem=524288 >>>>> >>>>> Test results: >>>>> SMC-R >>>>> msgsize noautosplit autosplit >>>>> 128KB 55.546 us 55.763 us >>>>> 256KB 83.537 us 69.743 us (17% improve) >>>>> 512KB 138.306 us 100.313 us (28% improve) >>>>> 1MB 273.702 us 197.222 us (28% improve) >>>>> >>>>> SMC-D with smc-loopback >>>>> msgsize noautosplit autosplit >>>>> 128KB 14.672 us 14.690 us >>>>> 256KB 28.277 us 23.958 us (15% improve) >>>>> 512KB 63.047 us 45.339 us (28% improve) >>>>> 1MB 129.306 us 87.278 us (32% improve) >>>>> >>>>> Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com> >>>>> --- >>>>> Documentation/networking/smc-sysctl.rst | 11 +++++++++++ >>>>> include/net/netns/smc.h | 1 + >>>>> net/smc/smc_sysctl.c | 12 ++++++++++++ >>>>> net/smc/smc_tx.c | 19 ++++++++++++++++++- >>>>> 4 files changed, 42 insertions(+), 1 deletion(-) >>>>> >>>> >>>> Hi Guangguan, >>>> >>>> If I remember correctly, the intention to use one RDMA-write for a possible large data is to reduce possible many partial stores. Since many year has gone, I'm not that sure if it would still be an issue. I need some time to check on it. >>>> >>> >>> Did you mean too many partial stores will result in some issue? What's the issue? >>> > Forget it, I did verify that the partial stores should not be problem now. >>> >>>> BTW, I don't really like the idea to use sysctl to set the autosplit_size in any value at will. That makes no sense to improve the performance. >>> >>> Although 128KB autosplit_size have a good performance in most scenario, I still found some better autosplit_size for some specific network configurations. >>> For example, 128KB autosplit_size have a good performance whether the MTU is 1500 or 8500, but for 8500 MTU, 64KB autosplit_size performs better. >>> >>> Maybe the sysctl is not the best way, but I think it should have a way to set the value of autosplit_size for possible performance tuning. >>> > mhhh, that could be a good reason to use sysctl. >>> Thanks, >>> Guangguan Wang >>> >> >> Hi Wenjia, >> >> Is there any update comment or information about this patch? > > Hi Guangguan, > > sorry for the delayed answer. In the last time it is really difficult for me to find time to look into it and test it. With more thinking, I'm kind of convinced with this idea. But test is still needed. I'll be in vacation next 3 weeks. I hope it is okay for you that I'll test it as soon as possible when I'm back. If everything is ok, I think we can let it go upstream. > > Thanks, > Wenjia Hi Wenjia, It is OK to test the patch after your vacation. Have a nice vacation. Thanks, Guangguan Wang. > > >> >>>> >>>> Thanks, >>>> Wenjia
diff --git a/Documentation/networking/smc-sysctl.rst b/Documentation/networking/smc-sysctl.rst index a874d007f2db..81b5296d79f4 100644 --- a/Documentation/networking/smc-sysctl.rst +++ b/Documentation/networking/smc-sysctl.rst @@ -71,3 +71,14 @@ smcr_max_conns_per_lgr - INTEGER acceptable value ranges from 16 to 255. Only for SMC-R v2.1 and later. Default: 255 + +autosplit_size - INTEGER + Setting SMC autosplit size. Autosplit is used to split sending data into + several segments when application sending data and the data size is larger + than autosplit size. Autosplit helps performing pipeline sending and pipeline + receiving for better latency performance when sending/receiving large size + data. + Autosplit_size ranges from 32KiB to 512MiB. Set autosplit_size to 512MiB means + disable autosplit. + + Default: 128KiB diff --git a/include/net/netns/smc.h b/include/net/netns/smc.h index fc752a50f91b..26c7edeb71a3 100644 --- a/include/net/netns/smc.h +++ b/include/net/netns/smc.h @@ -24,5 +24,6 @@ struct netns_smc { int sysctl_rmem; int sysctl_max_links_per_lgr; int sysctl_max_conns_per_lgr; + unsigned int sysctl_autosplit_size; }; #endif diff --git a/net/smc/smc_sysctl.c b/net/smc/smc_sysctl.c index 13f2bc092db1..2aaf402acc11 100644 --- a/net/smc/smc_sysctl.c +++ b/net/smc/smc_sysctl.c @@ -29,6 +29,8 @@ static int links_per_lgr_min = SMC_LINKS_ADD_LNK_MIN; static int links_per_lgr_max = SMC_LINKS_ADD_LNK_MAX; static int conns_per_lgr_min = SMC_CONN_PER_LGR_MIN; static int conns_per_lgr_max = SMC_CONN_PER_LGR_MAX; +static unsigned int autosplit_size_min = SZ_32K; +static unsigned int autosplit_size_max = SZ_512M; /* max size of snd/recv buffer */ static struct ctl_table smc_table[] = { { @@ -90,6 +92,15 @@ static struct ctl_table smc_table[] = { .extra1 = &conns_per_lgr_min, .extra2 = &conns_per_lgr_max, }, + { + .procname = "autosplit_size", + .data = &init_net.smc.sysctl_autosplit_size, + .maxlen = sizeof(unsigned int), + .mode = 0644, + .proc_handler = proc_douintvec_minmax, + .extra1 = &autosplit_size_min, + .extra2 = &autosplit_size_max, + }, }; int __net_init smc_sysctl_net_init(struct net *net) @@ -121,6 +132,7 @@ int __net_init smc_sysctl_net_init(struct net *net) WRITE_ONCE(net->smc.sysctl_rmem, net_smc_rmem_init); net->smc.sysctl_max_links_per_lgr = SMC_LINKS_PER_LGR_MAX_PREFER; net->smc.sysctl_max_conns_per_lgr = SMC_CONN_PER_LGR_PREFER; + net->smc.sysctl_autosplit_size = SZ_128K; return 0; diff --git a/net/smc/smc_tx.c b/net/smc/smc_tx.c index 214ac3cbcf9a..331ce4ff7c6e 100644 --- a/net/smc/smc_tx.c +++ b/net/smc/smc_tx.c @@ -175,6 +175,21 @@ static bool smc_tx_should_cork(struct smc_sock *smc, struct msghdr *msg) return false; } +static inline bool smc_tx_should_split(struct smc_sock *smc, size_t *len) +{ + size_t split_size = sock_net(&smc->sk)->smc.sysctl_autosplit_size; + + /* only split when len >= sysctl_autosplit_size * 1.3, + * in case of a following tiny size xmit. + */ + if (*len >= (split_size * 4 / 3)) { + *len = split_size; + return true; + } + + return false; +} + /* sndbuf producer: main API called by socket layer. * called under sock lock. */ @@ -185,6 +200,7 @@ int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len) struct smc_connection *conn = &smc->conn; union smc_host_cursor prep; struct sock *sk = &smc->sk; + bool is_split = false; char *sndbuf_base; int tx_cnt_prep; int writespace; @@ -235,6 +251,7 @@ int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len) writespace = atomic_read(&conn->sndbuf_space); /* not more than what user space asked for */ copylen = min_t(size_t, send_remaining, writespace); + is_split = smc_tx_should_split(smc, ©len); /* determine start of sndbuf */ sndbuf_base = conn->sndbuf_desc->cpu_addr; smc_curs_copy(&prep, &conn->tx_curs_prep, conn); @@ -281,7 +298,7 @@ int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len) /* If we need to cork, do nothing and wait for the next * sendmsg() call or push on tx completion */ - if (!smc_tx_should_cork(smc, msg)) + if (is_split || !smc_tx_should_cork(smc, msg)) smc_tx_sndbuf_nonempty(conn); trace_smc_tx_sendmsg(smc, copylen);