-----Original Message----- From: Song Bao Hua (Barry Song) Sent: Friday, March 19, 2021 5:16 PM To: tim.c.chen@linux.intel.com; catalin.marinas@arm.com; will@kernel.org; rjw@rjwysocki.net; vincent.guittot@linaro.org; bp@alien8.de; tglx@linutronix.de; mingo@redhat.com; lenb@kernel.org; peterz@infradead.org; dietmar.eggemann@arm.com; rostedt@goodmis.org; bsegall@google.com; mgorman@suse.de Cc: msys.mizuma@gmail.com; valentin.schneider@arm.com; gregkh@linuxfoundation.org; Jonathan Cameron jonathan.cameron@huawei.com; juri.lelli@redhat.com; mark.rutland@arm.com; sudeep.holla@arm.com; aubrey.li@linux.intel.com; linux-arm-kernel@lists.infradead.org; linux-kernel@vger.kernel.org; linux-acpi@vger.kernel.org; x86@kernel.org; xuwei (O) xuwei5@huawei.com; Zengtao (B) prime.zeng@hisilicon.com; guodong.xu@linaro.org; yangyicong yangyicong@huawei.com; Liguozhu (Kenneth) liguozhu@hisilicon.com; linuxarm@openeuler.org; hpa@zytor.com; Song Bao Hua (Barry Song) song.bao.hua@hisilicon.com Subject: [RFC PATCH v5 3/4] scheduler: scan idle cpu in cluster before scanning the whole llc
On kunpeng920, cpus within one cluster can communicate wit each other much faster than cpus across different clusters. A simple hackbench can prove that. hackbench running on 4 cpus in single one cluster and 4 cpus in different clusters shows a large contrast: (1) within a cluster: root@ubuntu:~# taskset -c 0,1,2,3 hackbench -p -T -l 20000 -g 1 Running in threaded mode with 1 groups using 40 file descriptors each (== 40 tasks) Each sender will pass 20000 messages of 100 bytes Time: 4.285
(2) across clusters: root@ubuntu:~# taskset -c 0,4,8,12 hackbench -p -T -l 20000 -g 1 Running in threaded mode with 1 groups using 40 file descriptors each (== 40 tasks) Each sender will pass 20000 messages of 100 bytes Time: 5.524
This inspires us to change the wake_affine path to scan cluster before scanning the whole LLC to try to gatter related tasks in one cluster, which is done by this patch.
To evaluate the performance impact to related tasks talking with each other, we run the below hackbench with different -g parameter from 2 to 14, for each different g, we run the command 10 times and get the average time: $ numactl -N 0 hackbench -p -T -l 20000 -g $1
hackbench will report the time which is needed to complete a certain number of messages transmissions between a certain number of tasks, for example: $ numactl -N 0 hackbench -p -T -l 20000 -g 10 Running in threaded mode with 10 groups using 40 file descriptors each (== 400 tasks) Each sender will pass 20000 messages of 100 bytes
The below is the result of hackbench w/ and w/o cluster patch: g= 2 4 6 8 10 12 14 w/o: 1.8151 3.8499 5.5142 7.2491 9.0340 10.7345 12.0929 w/ : 1.7881 3.7371 5.3301 6.9747 8.6909 9.9235 11.2608
Obviously some recent commits have improved the hackbench. So the change in wake_affine path brings less increase on hackbench compared to what we got in RFC v4. And obviously it is much more tricky to leverage wake_affine compared to leveraging the scatter of tasks in the previous patch as load balance might pull tasks which have been compact in a cluster so alternative suggestions welcome.
In order to figure out how many times cpu is picked from the cluster and how many times cpu is picked out of the cluster, a tracepoint for debug purpose is added in this patch. And an userspace bcc script to print the histogram of the result of select_idle_cpu(): #!/usr/bin/python # # selectidlecpu.py select idle cpu histogram. # # A Ctrl-C will print the gathered histogram then exit. # # 18-March-2021 Barry Song Created this.
from __future__ import print_function from bcc import BPF from time import sleep
# load BPF program b = BPF(text="""
BPF_HISTOGRAM(dist);
TRACEPOINT_PROBE(sched, sched_select_idle_cpu) { u32 e; if (args->idle / 4 == args->target/4) e = 0; /* idle cpu from same cluster */
Oops here, as -1/4 = 1/4 = 2/4 = 3/4 = 0 So a part of -1 is put here(local cluster) incorrectly.
else if (args->idle != -1) e = 1; /* idle cpu from different clusters */ else e = 2; /* no idle cpu */
dist.increment(e); return 0; } """)
Fixed it to:
TRACEPOINT_PROBE(sched, sched_select_idle_cpu) { u32 e; if (args->idle == -1) e = 2; /* no idle cpu */ else if (args->idle / 4 == args->target / 4) e = 0; /* idle cpu from same cluster */ else e = 1; /* idle cpu from different clusters */
dist.increment(e); return 0; }
# header print("Tracing... Hit Ctrl-C to end.")
# trace until Ctrl-C try: sleep(99999999) except KeyboardInterrupt: print()
# output
print("\nlinear histogram") print("~~~~~~~~~~~~~~~~") b["dist"].print_linear_hist("idle")
Even while g=14 and the system is quite busy, we can see there are some chances idle cpu is picked from local cluster: linear histogram
idle : count distribution 0 : 15234281 |*********** | 1 : 18494 | | 2 : 53066152 |****************************************| 0: local cluster 1: out of the cluster 2: select_idle_cpu() returns -1
After fixing the script, the new histogram is like: linear histogram ~~~~~~~~~~~~~~~~ idle : count distribution 0 : 2765930 |* | 1 : 68934 | | 2 : 77667475 |****************************************|
We get Local cluster: 3.4358% Out of cluster: 0.0856% -1(no idle before nr becomes 0): 96.4785%
Signed-off-by: Barry Song song.bao.hua@hisilicon.com
include/trace/events/sched.h | 22 ++++++++++++++++++++++ kernel/sched/fair.c | 32 +++++++++++++++++++++++++++++++- 2 files changed, 53 insertions(+), 1 deletion(-)
diff --git a/include/trace/events/sched.h b/include/trace/events/sched.h index cbe3e15..86608cf 100644 --- a/include/trace/events/sched.h +++ b/include/trace/events/sched.h @@ -136,6 +136,28 @@ );
/*
- Tracepoint for select_idle_cpu:
- */
+TRACE_EVENT(sched_select_idle_cpu,
- TP_PROTO(int target, int idle),
- TP_ARGS(target, idle),
- TP_STRUCT__entry(
__field( int, target )__field( int, idle )- ),
- TP_fast_assign(
__entry->target = target;__entry->idle = idle;- ),
- TP_printk("target=%d idle=%d", __entry->target, __entry->idle)
+);
+/*
- Tracepoint for waking up a task:
*/ DECLARE_EVENT_CLASS(sched_wakeup_template, diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index c92ad9f2..3892d42 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -6150,7 +6150,12 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int t if (!this_sd) return -1;
- cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
- if (!sched_cluster_active())
cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);+#ifdef CONFIG_SCHED_CLUSTER
- if (sched_cluster_active())
cpumask_and(cpus, cpu_cluster_mask(target), p->cpus_ptr);+#endif
if (sched_feat(SIS_PROP) && !smt) { u64 avg_cost, avg_idle, span_avg; @@ -6171,6 +6176,29 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int t time = cpu_clock(this); }
+#ifdef CONFIG_SCHED_CLUSTER
- if (sched_cluster_active()) {
for_each_cpu_wrap(cpu, cpus, target) {if (smt) {i = select_idle_core(p, cpu, cpus, &idle_cpu);if ((unsigned int)i < nr_cpumask_bits)return i;} else {if (!--nr)return -1;idle_cpu = __select_idle_cpu(cpu);if ((unsigned int)idle_cpu < nr_cpumask_bits) {goto done;}}
BTW, if I return -1 here directly and don't fall back to LLC, I can even get a better benchmark:
g= 2 4 6 8 10 12 14 w/o: 1.8151 3.8499 5.5142 7.2491 9.0340 10.7345 12.0929 w/ : 1.7881 3.7371 5.3301 6.9747 8.6909 9.9235 11.2608 return -1: 1.8324 3.6140 5.1029 6.5016 8.1867 9.7559 10.7716
so it seems the wake-up path change is much more trivial to get a real and good impact, comparing to the previous 2/4 patch in which we are only spreading tasks.
}cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);cpumask_andnot(cpus, cpus, cpu_cluster_mask(target));- }
+#endif
- for_each_cpu_wrap(cpu, cpus, target) { if (smt) { i = select_idle_core(p, cpu, cpus, &idle_cpu);
@@ -6186,6 +6214,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int t } }
+done: if (smt) set_idle_cores(this, false);
@@ -6324,6 +6353,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target) return target;
i = select_idle_cpu(p, sd, target);
- trace_sched_select_idle_cpu(target, i); if ((unsigned)i < nr_cpumask_bits) return i;
-- 1.8.3.1
Thanks Barry