[PATCH 1/1] sched/cputime: Mitigate performance regression in times()/clock_gettime()

From: Giovanni Gherdovich
Date: Fri Aug 05 2016 - 04:22:40 EST

Commit 6e998916dfe3 ("sched/cputime: Fix clock_nanosleep()/clock_gettime()
inconsistency") fixed a problem whereby clock_nanosleep() followed by
clock_gettime() could allow a task to wake early. It addressed the problem
by calling the scheduling classes update_curr when the cputimer starts.

Said change induced a considerable performance regression on the syscalls
times() and clock_gettimes(CLOCK_PROCESS_CPUTIME_ID). There are some
debuggers and applications that monitor their own performance that
accidentally depend on the performance of these specific calls.

This patch mitigates the performace loss by prefetching data in the CPU
cache, as stalls due to cache misses appear to be where most time is spent
in our benchmarks.

Here are the performance gain of this patch over v4.7-rc7 on a Sandy Bridge
box with 32 logical cores and 2 NUMA nodes. The test is repeated with a
variable number of threads, from 2 to 4*num_cpus; the results are in
seconds and correspond to the average of 10 runs; the percentage gain is
computed with (before-after)/before so a positive value is an improvement
(it's faster). The improvement varies between a few percents for 5-20
threads and more than 10% for 2 or >20 threads.

pound_clock_gettime:

threads 4.7-rc7 patched 4.7-rc7
[num] [secs] [secs (percent)]
2 3.48 3.06 ( 11.83%)
5 3.33 3.25 ( 2.40%)
8 3.37 3.26 ( 3.30%)
12 3.32 3.37 ( -1.60%)
21 4.01 3.90 ( 2.74%)
30 3.63 3.36 ( 7.41%)
48 3.71 3.11 ( 16.27%)
79 3.75 3.16 ( 15.74%)
110 3.81 3.25 ( 14.80%)
128 3.88 3.31 ( 14.76%)

pound_times:

threads 4.7-rc7 patched 4.7-rc7
[num] [secs] [secs (percent)]
2 3.65 3.25 ( 11.03%)
5 3.45 3.17 ( 7.92%)
8 3.52 3.22 ( 8.69%)
12 3.29 3.36 ( -2.04%)
21 4.07 3.92 ( 3.78%)
30 3.87 3.40 ( 12.17%)
48 3.79 3.16 ( 16.61%)
79 3.88 3.28 ( 15.42%)
110 3.90 3.38 ( 13.35%)
128 4.00 3.38 ( 15.45%)

pound_clock_gettime and pound_clock_gettime are two benchmarks included in
the MMTests framework. They launch a given number of threads which
repeatedly call times() or clock_gettimes(). The results above can be
reproduced with cloning MMTests from github.com and running the "poundtime"
workload:

$ git clone https://github.com/gormanm/mmtests.git
$ cd mmtests
$ cp configs/config-global-dhp__workload_poundtime config
$ ./run-mmtests.sh --run-monitor $(uname -r)

The above will run "poundtime" measuring the kernel currently running on
the machine; Once a new kernel is installed and the machine rebooted,
running again

$ cd mmtests
$ ./run-mmtests.sh --run-monitor $(uname -r)

will produce results to compare with. A comparison table will be output
with

$ cd mmtests/work/log
$ ../../compare-kernels.sh

the table will contain a lot of entries; grepping for "Amean" (as in
"arithmetic mean") will give the tables presented above. The source code
for the two benchmarks is reported at the end of this changelog for
clairity.

The cache misses addressed by this patch were found using a combination of
`perf top`, `perf record` and `perf annotate`. The incriminated lines were
found to be

struct sched_entity *curr = cfs_rq->curr;

and

delta_exec = now - curr->exec_start;

in the function update_curr() from kernel/sched/fair.c. This patch
prefetches the data from memory just before update_curr is called in the
interested execution path.

A comparison of the total number of cycles before and after the patch
follows; the data is obtained using `perf stat -r 10 -ddd <program>`
running over the same sequence of number of threads used above (a positive
gain is an improvement):

threads cycles before cycles after gain

2 19,699,563,964 +-1.19% 17,358,917,517 +-1.85% 11.88%
5 47,401,089,566 +-2.96% 45,103,730,829 +-0.97% 4.85%
8 80,923,501,004 +-3.01% 71,419,385,977 +-0.77% 11.74%
12 112,326,485,473 +-0.47% 110,371,524,403 +-0.47% 1.74%
21 193,455,574,299 +-0.72% 180,120,667,904 +-0.36% 6.89%
30 315,073,519,013 +-1.64% 271,222,225,950 +-1.29% 13.92%
48 321,969,515,332 +-1.48% 273,353,977,321 +-1.16% 15.10%
79 337,866,003,422 +-0.97% 289,462,481,538 +-1.05% 14.33%
110 338,712,691,920 +-0.78% 290,574,233,170 +-0.77% 14.21%
128 348,384,794,006 +-0.50% 292,691,648,206 +-0.66% 15.99%

A comparison of cache miss vs total cache loads ratios, before and after
the patch (again from the `perf stat -r 10 -ddd <program>` tables):

threads L1 misses/total*100 L1 misses/total*100 gain
before after
2 7.43 +-4.90% 7.36 +-4.70% 0.94%
5 13.09 +-4.74% 13.52 +-3.73% -3.28%
8 13.79 +-5.61% 12.90 +-3.27% 6.45%
12 11.57 +-2.44% 8.71 +-1.40% 24.72%
21 12.39 +-3.92% 9.97 +-1.84% 19.53%
30 13.91 +-2.53% 11.73 +-2.28% 15.67%
48 13.71 +-1.59% 12.32 +-1.97% 10.14%
79 14.44 +-0.66% 13.40 +-1.06% 7.20%
110 15.86 +-0.50% 14.46 +-0.59% 8.83%
128 16.51 +-0.32% 15.06 +-0.78% 8.78%

As a final note, the following shows the evolution of performance figures
in the "poundtime" benchmark and pinpoints commit 6e998916dfe3
("sched/cputime: Fix clock_nanosleep()/clock_gettime() inconsistency") as a
major source of degradation, mostly unaddressed to this day (figures
expressed in seconds).

pound_clock_gettime:

threads parent of 6e998916dfe3 4.7-rc7
6e998916dfe3 itself
2 2.23 3.68 ( -64.56%) 3.48 (-55.48%)
5 2.83 3.78 ( -33.42%) 3.33 (-17.43%)
8 2.84 4.31 ( -52.12%) 3.37 (-18.76%)
12 3.09 3.61 ( -16.74%) 3.32 ( -7.17%)
21 3.14 4.63 ( -47.36%) 4.01 (-27.71%)
30 3.28 5.75 ( -75.37%) 3.63 (-10.80%)
48 3.02 6.05 (-100.56%) 3.71 (-22.99%)
79 2.88 6.30 (-118.90%) 3.75 (-30.26%)
110 2.95 6.46 (-119.00%) 3.81 (-29.24%)
128 3.05 6.42 (-110.08%) 3.88 (-27.04%)

pound_times:

threads parent of 6e998916dfe3 4.7-rc7
6e998916dfe3 itself
2 2.27 3.73 ( -64.71%) 3.65 (-61.14%)
5 2.78 3.77 ( -35.56%) 3.45 (-23.98%)
8 2.79 4.41 ( -57.71%) 3.52 (-26.05%)
12 3.02 3.56 ( -17.94%) 3.29 ( -9.08%)
21 3.10 4.61 ( -48.74%) 4.07 (-31.34%)
30 3.33 5.75 ( -72.53%) 3.87 (-16.01%)
48 2.96 6.06 (-105.04%) 3.79 (-28.10%)
79 2.88 6.24 (-116.83%) 3.88 (-34.81%)
110 2.98 6.37 (-114.08%) 3.90 (-31.12%)
128 3.10 6.35 (-104.61%) 4.00 (-28.87%)

The source code of the two benchmarks follows. To compile the two:

NR_THREADS=42
for FILE in pound_times pound_clock_gettime; do
gcc -lrt -O2 -lpthread -DNUM_THREADS=$NR_THREADS $FILE.c -o $FILE
done

==== BEGIN pound_times.c ====

struct tms start;

void *pound (void *threadid)
{
struct tms end;
int oldutime = 0;
int utime;
int i;
for (i = 0; i < 5000000 / NUM_THREADS; i++) {
times(&end);
utime = ((int)end.tms_utime - (int)start.tms_utime);
if (oldutime > utime) {
printf("utime decreased, was %d, now %d!\n", oldutime, utime);
}
oldutime = utime;
}
pthread_exit(NULL);
}

int main()
{
pthread_t th[NUM_THREADS];
long i;
times(&start);
for (i = 0; i < NUM_THREADS; i++) {
pthread_create (&th[i], NULL, pound, (void *)i);
}
pthread_exit(NULL);
return 0;
}
==== END pound_times.c ====

==== BEGIN pound_clock_gettime.c ====

void *pound (void *threadid)
{
struct timespec ts;
int rc, i;
unsigned long prev = 0, this = 0;

for (i = 0; i < 5000000 / NUM_THREADS; i++) {
rc = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
if (rc < 0)
perror("clock_gettime");
this = (ts.tv_sec * 1000000000) + ts.tv_nsec;
if (0 && this < prev)
printf("%lu ns timewarp at iteration %d\n", prev - this, i);
prev = this;
}
pthread_exit(NULL);
}

int main()
{
pthread_t th[NUM_THREADS];
long rc, i;
pid_t pgid;

for (i = 0; i < NUM_THREADS; i++) {
rc = pthread_create(&th[i], NULL, pound, (void *)i);
if (rc < 0)
perror("pthread_create");
}

pthread_exit(NULL);
return 0;
}
==== END pound_clock_gettime.c ====

Signed-off-by: Giovanni Gherdovich <ggherdovich@xxxxxxx>
Suggested-by: Mike Galbraith <mgalbraith@xxxxxxx>
---
kernel/sched/core.c | 19 +++++++++++++++++++
1 file changed, 19 insertions(+)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 51d7105..4500421 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -74,6 +74,7 @@
#include <linux/context_tracking.h>
#include <linux/compiler.h>
#include <linux/frame.h>
+#include <linux/prefetch.h>

#include <asm/switch_to.h>
#include <asm/tlb.h>
@@ -2965,6 +2966,23 @@ EXPORT_PER_CPU_SYMBOL(kstat);
EXPORT_PER_CPU_SYMBOL(kernel_cpustat);

/*
+ * The function fair_sched_class.update_curr accesses the struct curr
+ * and its field curr->exec_start; when called from task_sched_runtime,
+ * we observe a high rate of cache misses in practice.
+ * Prefetching this data results in improved performance.
+ */
+static inline void prefetch_curr_exec_start(struct task_struct *p)
+{
+#ifdef CONFIG_FAIR_GROUP_SCHED
+ struct sched_entity *curr = (&p->se)->cfs_rq->curr;
+#else
+ struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
+#endif
+ prefetch(curr);
+ prefetch(&curr->exec_start);
+}
+
+/*
* Return accounted runtime for the task.
* In case the task is currently running, return the runtime plus current's
* pending runtime that have not been accounted yet.
@@ -2998,6 +3016,7 @@ unsigned long long task_sched_runtime(struct task_struct *p)
* thread, breaking clock_gettime().
*/
if (task_current(rq, p) && task_on_rq_queued(p)) {
+ prefetch_curr_exec_start(p);
update_rq_clock(rq);
p->sched_class->update_curr(rq);
}
--
2.6.6