From f8858d96061f5942216c6abb0194c3ea7b78e1e8 Mon Sep 17 00:00:00 2001 From: Shrikanth Hegde Date: Sat, 2 Sep 2023 13:42:04 +0530 Subject: sched/fair: Optimize should_we_balance() for large SMT systems MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit should_we_balance() is called in load_balance() to find out if the CPU that is trying to do the load balance is the right one or not. With commit: b1bfeab9b002("sched/fair: Consider the idle state of the whole core for load balance") the code tries to find an idle core to do the load balancing and falls back on an idle sibling CPU if there is no idle core. However, on larger SMT systems, it could be needlessly iterating to find a idle by scanning all the CPUs in an non-idle core. If the core is not idle, and first SMT sibling which is idle has been found, then its not needed to check other SMT siblings for idleness Lets say in SMT4, Core0 has 0,2,4,6 and CPU0 is BUSY and rest are IDLE. balancing domain is MC/DIE. CPU2 will be set as the first idle_smt and same process would be repeated for CPU4 and CPU6 but this is unnecessary. Since calling is_core_idle loops through all CPU's in the SMT mask, effect is multiplied by weight of smt_mask. For example,when say 1 CPU is busy, we would skip loop for 2 CPU's and skip iterating over 8CPU's. That effect would be more in DIE/NUMA domain where there are more cores. Testing and performance evaluation ================================== The test has been done on this system which has 12 cores, i.e 24 small cores with SMT=4: lscpu Architecture: ppc64le Byte Order: Little Endian CPU(s): 96 On-line CPU(s) list: 0-95 Model name: POWER10 (architected), altivec supported Thread(s) per core: 8 Used funclatency bcc tool to evaluate the time taken by should_we_balance(). For base tip/sched/core the time taken is collected by making the should_we_balance() noinline. time is in nanoseconds. The values are collected by running the funclatency tracer for 60 seconds. values are average of 3 such runs. This represents the expected reduced time with patch. tip/sched/core was at commit: 2f88c8e802c8 ("sched/eevdf/doc: Modify the documented knob to base_slice_ns as well") Results: ------------------------------------------------------------------------------ workload tip/sched/core with_patch(%gain) ------------------------------------------------------------------------------ idle system 809.3 695.0(16.45) stress ng – 12 threads -l 100 1013.5 893.1(13.49) stress ng – 24 threads -l 100 1073.5 980.0(9.54) stress ng – 48 threads -l 100 683.0 641.0(6.55) stress ng – 96 threads -l 100 2421.0 2300(5.26) stress ng – 96 threads -l 15 375.5 377.5(-0.53) stress ng – 96 threads -l 25 635.5 637.5(-0.31) stress ng – 96 threads -l 35 934.0 891.0(4.83) Ran schbench(old), hackbench and stress_ng to evaluate the workload performance between tip/sched/core and with patch. No modification to tip/sched/core TL;DR: Good improvement is seen with schbench. when hackbench and stress_ng runs for longer good improvement is seen. ------------------------------------------------------------------------------ schbench(old) tip +patch(%gain) 10 iterations sched/core ------------------------------------------------------------------------------ 1 Threads 50.0th: 8.00 9.00(-12.50) 75.0th: 9.60 9.00(6.25) 90.0th: 11.80 10.20(13.56) 95.0th: 12.60 10.40(17.46) 99.0th: 13.60 11.90(12.50) 99.5th: 14.10 12.60(10.64) 99.9th: 15.90 14.60(8.18) 2 Threads 50.0th: 9.90 9.20(7.07) 75.0th: 12.60 10.10(19.84) 90.0th: 15.50 12.00(22.58) 95.0th: 17.70 14.00(20.90) 99.0th: 21.20 16.90(20.28) 99.5th: 22.60 17.50(22.57) 99.9th: 30.40 19.40(36.18) 4 Threads 50.0th: 12.50 10.60(15.20) 75.0th: 15.30 12.00(21.57) 90.0th: 18.60 14.10(24.19) 95.0th: 21.30 16.20(23.94) 99.0th: 26.00 20.70(20.38) 99.5th: 27.60 22.50(18.48) 99.9th: 33.90 31.40(7.37) 8 Threads 50.0th: 16.30 14.30(12.27) 75.0th: 20.20 17.40(13.86) 90.0th: 24.50 21.90(10.61) 95.0th: 27.30 24.70(9.52) 99.0th: 35.00 31.20(10.86) 99.5th: 46.40 33.30(28.23) 99.9th: 89.30 57.50(35.61) 16 Threads 50.0th: 22.70 20.70(8.81) 75.0th: 30.10 27.40(8.97) 90.0th: 36.00 32.80(8.89) 95.0th: 39.60 36.40(8.08) 99.0th: 49.20 44.10(10.37) 99.5th: 64.90 50.50(22.19) 99.9th: 143.50 100.60(29.90) 32 Threads 50.0th: 34.60 35.50(-2.60) 75.0th: 48.20 50.50(-4.77) 90.0th: 59.20 62.40(-5.41) 95.0th: 65.20 69.00(-5.83) 99.0th: 80.40 83.80(-4.23) 99.5th: 102.10 98.90(3.13) 99.9th: 727.10 506.80(30.30) schbench does improve in general. There is some run to run variation with schbench. Did a validation run to confirm that trend is similar. ------------------------------------------------------------------------------ hackbench tip +patch(%gain) 20 iterations, 50000 loops sched/core ------------------------------------------------------------------------------ Process 10 groups : 11.74 11.70(0.34) Process 20 groups : 22.73 22.69(0.18) Process 30 groups : 33.39 33.40(-0.03) Process 40 groups : 43.73 43.61(0.27) Process 50 groups : 53.82 54.35(-0.98) Process 60 groups : 64.16 65.29(-1.76) thread 10 Time : 12.81 12.79(0.16) thread 20 Time : 24.63 24.47(0.65) Process(Pipe) 10 Time : 6.40 6.34(0.94) Process(Pipe) 20 Time : 10.62 10.63(-0.09) Process(Pipe) 30 Time : 15.09 14.84(1.66) Process(Pipe) 40 Time : 19.42 19.01(2.11) Process(Pipe) 50 Time : 24.04 23.34(2.91) Process(Pipe) 60 Time : 28.94 27.51(4.94) thread(Pipe) 10 Time : 6.96 6.87(1.29) thread(Pipe) 20 Time : 11.74 11.73(0.09) hackbench shows slight improvement with pipe. Slight degradation in process. ------------------------------------------------------------------------------ stress_ng tip +patch(%gain) 10 iterations 100000 cpu_ops sched/core ------------------------------------------------------------------------------ --cpu=96 -util=100 Time taken : 5.30, 5.01(5.47) --cpu=48 -util=100 Time taken : 7.94, 6.73(15.24) --cpu=24 -util=100 Time taken : 11.67, 8.75(25.02) --cpu=12 -util=100 Time taken : 15.71, 15.02(4.39) --cpu=96 -util=10 Time taken : 22.71, 22.19(2.29) --cpu=96 -util=20 Time taken : 12.14, 12.37(-1.89) --cpu=96 -util=30 Time taken : 8.76, 8.86(-1.14) --cpu=96 -util=40 Time taken : 7.13, 7.14(-0.14) --cpu=96 -util=50 Time taken : 6.10, 6.13(-0.49) --cpu=96 -util=60 Time taken : 5.42, 5.41(0.18) --cpu=96 -util=70 Time taken : 4.94, 4.94(0.00) --cpu=96 -util=80 Time taken : 4.56, 4.53(0.66) --cpu=96 -util=90 Time taken : 4.27, 4.26(0.23) Good improvement seen with 24 CPUs. In this case only one CPU is busy, and no core is idle. Decent improvement with 100% utilization case. no difference in other utilization. Fixes: b1bfeab9b002 ("sched/fair: Consider the idle state of the whole core for load balance") Signed-off-by: Shrikanth Hegde Signed-off-by: Ingo Molnar Link: https://lore.kernel.org/r/20230902081204.232218-1-sshegde@linux.vnet.ibm.com --- kernel/sched/fair.c | 15 ++++++++++++++- 1 file changed, 14 insertions(+), 1 deletion(-) (limited to 'kernel') diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 8dbff6e7ad4f..33a2b6bba676 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -6619,6 +6619,7 @@ dequeue_throttle: /* Working cpumask for: load_balance, load_balance_newidle. */ static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask); +static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask); #ifdef CONFIG_NO_HZ_COMMON @@ -10917,6 +10918,7 @@ static int active_load_balance_cpu_stop(void *data); static int should_we_balance(struct lb_env *env) { + struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask); struct sched_group *sg = env->sd->groups; int cpu, idle_smt = -1; @@ -10940,8 +10942,9 @@ static int should_we_balance(struct lb_env *env) return 1; } + cpumask_copy(swb_cpus, group_balance_mask(sg)); /* Try to find first idle CPU */ - for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { + for_each_cpu_and(cpu, swb_cpus, env->cpus) { if (!idle_cpu(cpu)) continue; @@ -10953,6 +10956,14 @@ static int should_we_balance(struct lb_env *env) if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) { if (idle_smt == -1) idle_smt = cpu; + /* + * If the core is not idle, and first SMT sibling which is + * idle has been found, then its not needed to check other + * SMT siblings for idleness: + */ +#ifdef CONFIG_SCHED_SMT + cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu)); +#endif continue; } @@ -12918,6 +12929,8 @@ __init void init_sched_fair_class(void) for_each_possible_cpu(i) { zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i)); zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i)); + zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i), + GFP_KERNEL, cpu_to_node(i)); #ifdef CONFIG_CFS_BANDWIDTH INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i)); -- cgit v1.2.3