Hi Barry and Tim,
Some more updates for the speccpu. Most results are rather positive and as expected.
For the mcf_r alone with numa 0 bound (5 iterations included):
[w/o] min max aver stddev 4 Copies 10.7 11.2 11 0.187082869 8 Copies 19.4 20.1 19.7 0.250998008 16 Copies 30.1 30.7 30.5 0.240831892
[w] min max aver stddev aver. enhancement 4 Copies 11.3 11.4 11.3 0.054772256 2.73% 8 Copies 21.1 21.4 21.3 0.114017543 8.12% 16 Copies 30.5 30.8 30.7 0.130384048 0.66%
8 Copies is the best case, as there are 8 clusters per NUMA and the threads can spread through the clusters well. For 4, 8, 16 copies we have less bouncing with cluster scheduler level, as the standard deviation is smaller.
For all intrate suite without numa bound and with 32 copies (4 iterations included):
[w/o] Min Max Aver Stddev 500.perlbench_r 86 87.1 86.7 0.496655481 502.gcc_r 89.9 91 90.25 0.519615242 505.mcf_r 71.5 73.6 72.375 0.970824392 520.omnetpp_r 40.3 41.2 40.6 0.40824829 523.xalancbmk_r 57.4 59.2 58.1 0.836660027 525.x264_r 197 198 197.5 0.577350269 531.deepsjeng_r 109 110 109.25 0.5 541.leela_r 95.4 95.6 95.525 0.095742711 548.exchange2_r 163 163 163 0 557.xz_r 64.2 65.2 64.85 0.450924975
Est. SPECrate2017_int_base 88.1
[w] Min Max Aver Stddev aver. Enhancement 500.perlbench_r 87.3 87.9 87.625 0.2500 1.07% 502.gcc_r 93.1 95.3 94.5 0.9933 4.71% 505.mcf_r 77 81.4 79.075 2.3514 9.26% 520.omnetpp_r 42.8 43.5 43.15 0.3109 6.28% 523.xalancbmk_r 60.6 62 61.45 0.6455 5.77% 525.x264_r 197 198 197.75 0.5000 0.13% 531.deepsjeng_r 109 110 109.75 0.5000 0.46% 541.leela_r 95.5 95.6 95.55 0.0577 0.03% 548.exchange2_r 163 163 163 0.0000 0.00% 557.xz_r 66 66.5 66.3 0.2160 2.24%
Est. SPECrate2017_int_base 90.7 (+2.95%)
The mcf_r performs even better. Although the stddev is larger than w/o result, but the minimum rate(77) is bigger than the maximum rate(73.6) of w/o result.
Well some benchmarks are not affected by the patch, I guess they'are cpu bound. We cannot decrease increase the bandwidth by place these threads inter-clusters.
The test machines are both 2 socket with 128 cores, 32 cores per numa and 4 cores per cluster.
Thanks, Yicong
On 2021/6/18 18:27, Yicong Yang wrote:
Hi Barry and Tim,
As Barry pointed I didn't enable the CONFIG_SCHED_CLUSTER... I'd like to share some updated results with the config correctly enabled.
I re-run the mcf_r with 4,8,16 copies on NUMA 0, the result is like: Base Base Run Time Rate ------- --------- 4 Copies w/o 580 (w 570) w/o 11.1 (w 11.3) 8 Copies w/o 647 (w 605) w/o 20.0 (w 21.4) 16 Copies w/o 844 (w 844) w/o 30.6 (w 30.6)
Seems there is a ~7% improvement for 8 Copies but little changed for 4 and 16 copies.
This time from htop the tasks are spread through clusters well.
For the 4 copies I use
perf stat -e probe:active_load_balance_cpu_stop -- ./bin/runcpu default.cfg 505.mcf_r
to check the different of 'active_load_balance_cpu_stop' with and without the patch. There is no difference and the counts are both 0.
I also run the whole intrate suite on another machine, same model as the one above. 32 Copies is lanuch in the whole system without binding to a specific NUMA node. And seems there is some positive results.
(x264_r is not included as there is a bug for x264 while compiling with gcc 10: https://www.spec.org/cpu2017/Docs/benchmarks/625.x264_s.html I'll fix this in the following test)
[w/o] Base Base Base Benchmarks Copies Run Time Rate
500.perlbench_r 32 584 87.2 * 502.gcc_r 32 503 90.2 * 505.mcf_r 32 745 69.4 * 520.omnetpp_r 32 1031 40.7 * 523.xalancbmk_r 32 597 56.6 * 525.x264_r 1 -- CE 531.deepsjeng_r 32 336 109 * 541.leela_r 32 556 95.4 * 548.exchange2_r 32 513 163 * 557.xz_r 32 530 65.2 * Est. SPECrate2017_int_base 80.3
[w] Base Base Base Benchmarks Copies Run Time Rate
500.perlbench_r 32 580 87.8 (+0.688%) * 502.gcc_r 32 477 95.1 (+5.432%) * 505.mcf_r 32 644 80.3 (+13.574%) * 520.omnetpp_r 32 942 44.6 (+9.58%) * 523.xalancbmk_r 32 560 60.4 (+6.714%%) * 525.x264_r 1 -- CE 531.deepsjeng_r 32 337 109 (+0.000%) * 541.leela_r 32 554 95.6 (+0.210%) * 548.exchange2_r 32 515 163 (+0.000%) * 557.xz_r 32 524 66.0 (+1.227%) * Est. SPECrate2017_int_base 83.7 (+4.062%)
The iteration of the test is 1, and I'm going to increase it to 5 to see the average result.
Thanks, Yicong
On 2021/6/17 19:33, Yicong Yang wrote:
On 2021/6/16 17:36, Barry Song wrote:
ARM64 chip Kunpeng 920 has 6 or 8 clusters in each NUMA node, and each cluster has 4 cpus. All clusters share L3 cache data, but each cluster has local L3 tag. On the other hand, each clusters will share some internal system bus. This means cache coherence overhead inside one cluster is much less than the overhead across clusters.
This patch adds the sched_domain for clusters. On kunpeng 920, without this patch, domain0 of cpu0 would be MC with cpu0~cpu23 with ; with this patch, MC becomes domain1, a new domain0 "CLS" including cpu0-cpu3.
This will help spread tasks among clusters, thus decrease the contention and improve the throughput. Verified by Mel's mm-tests by writing config files as below to define the number of threads for stream,e.g configs/config-workload-stream-omp-4threads: export STREAM_SIZE=$((1048576*512)) export STREAM_THREADS=4 export STREAM_METHOD=omp export STREAM_ITERATIONS=5 export STREAM_BUILD_FLAGS="-lm -Ofast"
Ran the stream benchmark on kunpeng920 with 4numa nodes and each node has 24core by commands like: numactl -N 0 -m 0 ./run-mmtests.sh -c \ configs/config-workload-stream-omp-4threads tip-sched-core-4threads
and compared the cases between tip/sched/core and tip/sched/core with cluster scheduler. The result is as below:
4threads stream (on 1numa * 24cores = 24cores) stream stream 4threads 4threads-cluster-scheduler MB/sec copy 29929.64 ( 0.00%) 32932.68 ( 10.03%) MB/sec scale 29861.10 ( 0.00%) 32710.58 ( 9.54%) MB/sec add 27034.42 ( 0.00%) 32400.68 ( 19.85%) MB/sec triad 27225.26 ( 0.00%) 31965.36 ( 17.41%)
6threads stream (on 1numa * 24cores = 24cores) stream stream 6threads 6threads-cluster-scheduler MB/sec copy 40330.24 ( 0.00%) 42377.68 ( 5.08%) MB/sec scale 40196.42 ( 0.00%) 42197.90 ( 4.98%) MB/sec add 37427.00 ( 0.00%) 41960.78 ( 12.11%) MB/sec triad 37841.36 ( 0.00%) 42513.64 ( 12.35%)
12threads stream (on 1numa * 24cores = 24cores) stream stream 12threads 12threads-cluster-scheduler MB/sec copy 52639.82 ( 0.00%) 53818.04 ( 2.24%) MB/sec scale 52350.30 ( 0.00%) 53253.38 ( 1.73%) MB/sec add 53607.68 ( 0.00%) 55198.82 ( 2.97%) MB/sec triad 54776.66 ( 0.00%) 56360.40 ( 2.89%)
The result was generated by commands like: ../../compare-kernels.sh --baseline tip-sched-core-4threads \ --compare tip-sched-core-4threads-cluster-scheduler
Thus, it could help memory-bound workload especially under medium load. For example, ran mmtests configs/config-workload-lkp-compress benchmark on 4numa*24cores=96 cores kunpeng920, 12,21,30 threads present the best improvement:
lkp-pbzip2 (on 4numa * 24cores = 96cores)
lkp lkp compress-w/o-cluster compress-w/-clusterHmean tput-2 11062841.57 ( 0.00%) 11341817.51 * 2.52%* Hmean tput-5 26815503.70 ( 0.00%) 27412872.65 * 2.23%* Hmean tput-8 41873782.21 ( 0.00%) 43326212.92 * 3.47%* Hmean tput-12 61875980.48 ( 0.00%) 64578337.51 * 4.37%* Hmean tput-21 105814963.07 ( 0.00%) 111381851.01 * 5.26%* Hmean tput-30 150349470.98 ( 0.00%) 156507070.73 * 4.10%* Hmean tput-48 237195937.69 ( 0.00%) 242353597.17 * 2.17%* Hmean tput-79 360252509.37 ( 0.00%) 362635169.23 * 0.66%* Hmean tput-96 394571737.90 ( 0.00%) 400952978.48 * 1.62%*
Ran the same benchmark by "numactl -N 0 -m 0" from 2 threads to 24 threads on numa node0 with 24 cores:
lkp-pbzip2 (on 1numa * 24cores = 24cores) lkp lkp compress-1numa-w/o-cluster compress-1numa-w/-cluster Hmean tput-2 11071705.49 ( 0.00%) 11296869.10 * 2.03%* Hmean tput-4 20782165.19 ( 0.00%) 21949232.15 * 5.62%* Hmean tput-6 30489565.14 ( 0.00%) 33023026.96 * 8.31%* Hmean tput-8 40376495.80 ( 0.00%) 42779286.27 * 5.95%* Hmean tput-12 61264033.85 ( 0.00%) 62995632.78 * 2.83%* Hmean tput-18 86697139.39 ( 0.00%) 86461545.74 ( -0.27%) Hmean tput-24 104854637.04 ( 0.00%) 104522649.46 * -0.32%*
In the case of 6 threads and 8 threads, we see the greatest performance improvement.
Similar improvement was seen on lkp-pixz though the improvement is smaller:
lkp-pixz (on 1numa * 24cores = 24cores) lkp lkp compress-1numa-w/o-cluster compress-1numa-w/-cluster Hmean tput-2 6486981.16 ( 0.00%) 6561515.98 * 1.15%* Hmean tput-4 11645766.38 ( 0.00%) 11614628.43 ( -0.27%) Hmean tput-6 15429943.96 ( 0.00%) 15957350.76 * 3.42%* Hmean tput-8 19974087.63 ( 0.00%) 20413746.98 * 2.20%* Hmean tput-12 28172068.18 ( 0.00%) 28751997.06 * 2.06%* Hmean tput-18 39413409.54 ( 0.00%) 39896830.55 * 1.23%* Hmean tput-24 49101815.85 ( 0.00%) 49418141.47 * 0.64%*
On the other hand, it is slightly helpful to cpu-bound tasks. With configs/config-workload-kernbench like: export KERNBENCH_ITERATIONS=3 export KERNBENCH_MIN_THREADS=$((NUMCPUS/4)) export KERNBENCH_MAX_THREADS=$((NUMCPUS)) export KERNBENCH_CONFIG=allmodconfig export KERNBENCH_TARGETS=vmlinux,modules export KERNBENCH_SKIP_WARMUP=yes export MMTESTS_THREAD_CUTOFF= export KERNBENCH_VERSION=5.3 Ran kernbench by 24,48,96 threads to compile an entire kernel without numactl binding, each case run 3 iterations: 24 threads w/o and w/ cluster-scheduler: w/o 10:03.26 10:00.46 10:01.09 w/ 10:01.11 10:00.83 9:58.64
48 threads w/o and w/ cluster-scheduler: w/o 5:33.96 5:34.28 5:34.06 w/ 5:32.65 5:32.57 5:33.25
96 threads w/o and w/ cluster-scheduler: w/o 3:33.34 3:31.22 3:31.31 w/ 3:32.22 3:30.47 3:32.69
kernbench (on 4numa * 24cores = 96cores) kernbench kernbench w/o-cluster w/-cluster Min user-24 12054.67 ( 0.00%) 12024.19 ( 0.25%) Min syst-24 1751.51 ( 0.00%) 1731.68 ( 1.13%) Min elsp-24 600.46 ( 0.00%) 598.64 ( 0.30%) Min user-48 12361.93 ( 0.00%) 12315.32 ( 0.38%) Min syst-48 1917.66 ( 0.00%) 1892.73 ( 1.30%) Min elsp-48 333.96 ( 0.00%) 332.57 ( 0.42%) Min user-96 12922.40 ( 0.00%) 12921.17 ( 0.01%) Min syst-96 2143.94 ( 0.00%) 2110.39 ( 1.56%) Min elsp-96 211.22 ( 0.00%) 210.47 ( 0.36%) Amean user-24 12063.99 ( 0.00%) 12030.78 * 0.28%* Amean syst-24 1755.20 ( 0.00%) 1735.53 * 1.12%* Amean elsp-24 601.60 ( 0.00%) 600.19 ( 0.23%) Amean user-48 12362.62 ( 0.00%) 12315.56 * 0.38%* Amean syst-48 1921.59 ( 0.00%) 1894.95 * 1.39%* Amean elsp-48 334.10 ( 0.00%) 332.82 * 0.38%* Amean user-96 12925.27 ( 0.00%) 12922.63 ( 0.02%) Amean syst-96 2146.66 ( 0.00%) 2122.20 * 1.14%* Amean elsp-96 211.96 ( 0.00%) 211.79 ( 0.08%)
[ Hi Yicong, Is it possible for you to run similar SPECrate mcf test with Tim and get some supportive data here?
For our Kunpeng 920, I run the whole intrate suite firstly with 32 copies, and didn't bind the NUMA. Here's the result
Base Base BaseBenchmarks Copies Run Time Rate
500.perlbench_r 32 w/o 580(w 578) w/o 87.8(w 88.2) * 502.gcc_r 32 w/o 500(w 504) w/o 90.5(w 90.0) * 505.mcf_r 32 w/o 764(w 767) w/o 67.7(w 67.4) * 520.omnetpp_r 32 w/o 1030(w 1024) w/o 40.7(w 41.0) * 523.xalancbmk_r 32 w/o 584(w 584) w/o 57.9(w 57.8) * 525.x264_r 32 w/o 285(w 284) w/o 196 (w 197) * 531.deepsjeng_r 32 w/o 336(w 338) w/o 109 (w 108) * 541.leela_r 32 w/o 570(w 569) w/o 93.0(w 93.1) * 548.exchange2_r 32 w/o 526(w 532) w/o 160 (w 157) * 557.xz_r 32 w/o 538(w 542) w/o 64.2(w 63.8) * Est. SPECrate2017_int_base w/o 87.4(w 87.2)
(w/o is without the patch, the bigger the rate is the better)
Then I test the mcf_r alone with different copies and bind to NUMA 0:
Base Base Run Time Rate ------- ---------4 Copies w/o 618 (w 580) w/o 10.5 (w 11.1) 8 Copies w/o 645 (w 647) w/o 20.0 (w 20) 16 Copies w/o 849 (w 844) w/o 30.4 (w 30.6)
As I checked from the htop, the tasks running on the cpu didn't spread through the clusters rigidly.
I didn't apply Patch #3 as I met some conflicts and didn't try to resolve it. As we're testing on arm64 I think it's okay to test without patch #3.
The machine I have tested have 128 cores in 2 sockets and 4 numas with 32 cores each. Of course, still 4 cores in one cluster. Below are the memory info through numa:
available: 4 nodes (0-3) node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 node 0 size: 257190 MB node 0 free: 254203 MB node 1 cpus: 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 node 1 size: 258005 MB node 1 free: 257191 MB node 2 cpus: 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 node 2 size: 96763 MB node 2 free: 96158 MB node 3 cpus: 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 node 3 size: 127540 MB node 3 free: 126922 MB node distances: node 0 1 2 3 0: 10 12 20 22 1: 12 10 22 24 2: 20 22 10 12 3: 22 24 12 10Any comments? I notice Tim observed that sleep and wakeup will have some influences. So I wonder whether the speccpu intrate test also suffers from this.
Thanks, Yicong
Thanks Barry ]
Signed-off-by: Barry Song song.bao.hua@hisilicon.com
arch/arm64/Kconfig | 7 +++++++ include/linux/sched/topology.h | 7 +++++++ include/linux/topology.h | 7 +++++++ kernel/sched/topology.c | 5 +++++ 4 files changed, 26 insertions(+)
diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig index 9f1d8566bbf9..3b54ea4e1bd7 100644 --- a/arch/arm64/Kconfig +++ b/arch/arm64/Kconfig @@ -999,6 +999,13 @@ config SCHED_MC making when dealing with multi-core CPU chips at a cost of slightly increased overhead in some places. If unsure say N here. +config SCHED_CLUSTER
- bool "Cluster scheduler support"
- help
Cluster scheduler support improves the CPU scheduler's decisionmaking when dealing with machines that have clusters(sharing internalbus or sharing LLC cache tag). If unsure say N here.config SCHED_SMT bool "SMT scheduler support" help diff --git a/include/linux/sched/topology.h b/include/linux/sched/topology.h index 8f0f778b7c91..2f9166f6dec8 100644 --- a/include/linux/sched/topology.h +++ b/include/linux/sched/topology.h @@ -42,6 +42,13 @@ static inline int cpu_smt_flags(void) } #endif +#ifdef CONFIG_SCHED_CLUSTER +static inline int cpu_cluster_flags(void) +{
- return SD_SHARE_PKG_RESOURCES;
+} +#endif
#ifdef CONFIG_SCHED_MC static inline int cpu_core_flags(void) { diff --git a/include/linux/topology.h b/include/linux/topology.h index 80d27d717631..0b3704ad13c8 100644 --- a/include/linux/topology.h +++ b/include/linux/topology.h @@ -212,6 +212,13 @@ static inline const struct cpumask *cpu_smt_mask(int cpu) } #endif +#if defined(CONFIG_SCHED_CLUSTER) && !defined(cpu_cluster_mask) +static inline const struct cpumask *cpu_cluster_mask(int cpu) +{
- return topology_cluster_cpumask(cpu);
+} +#endif
static inline const struct cpumask *cpu_cpu_mask(int cpu) { return cpumask_of_node(cpu_to_node(cpu)); diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c index 55a0a243e871..c7523dc7aab7 100644 --- a/kernel/sched/topology.c +++ b/kernel/sched/topology.c @@ -1511,6 +1511,11 @@ static struct sched_domain_topology_level default_topology[] = { #ifdef CONFIG_SCHED_SMT { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, #endif
+#ifdef CONFIG_SCHED_CLUSTER
{ cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },+#endif
#ifdef CONFIG_SCHED_MC { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, #endif
.
.