perf test: Increase load in lock contention test on low-activity systems
RFC PATCH 1/1 perf test: Increase load in lock contention test on low-activity systems
On low-activity systems, the 'kernel lock contention analysis test' often fails due to insufficient system load. To address this, the test now increases load by using multiple groups and threads in 'perf bench sched messaging', scaled to the number of available CPUs.
--- tools/perf/tests/shell/lock_contention.sh | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/tests/shell/lock_contention.sh +2 -1
@@ -44,7 +44,8 @@ check() { test_record() { echo "Testing perf lock record and perf lock contention" - perf lock record -o ${perfdata} -- perf bench sched messaging > /dev/null 2>&1 + perf lock record -o ${perfdata} -- perf bench sched messaging \ + --group "$(nproc)" --thread "$(nproc)" > /dev/null 2>&1 # the output goes to the stderr and we expect only 1 output (-E 1) perf lock contention -i ${perfdata} -E 1 -q 2> ${result} if [ "$(cat "${result}" | wc -l)" != "1" ]; then
--
2.48.1
Quoted 25 lines · Jan Polensky
On low-activity systems, the 'kernel lock contention analysis test' often fails due to insufficient system load. To address this, the test now increases load by using multiple groups and threads in 'perf bench sched messaging', scaled to the number of available CPUs.
Signed-off-by Jan Polensky <[email protected]>--- tools/perf/tests/shell/lock_contention.sh | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)tools/perf/tests/shell/lock_contention.sh +2
-2@@ -44,7 +44,8 @@ check() { test_record() { echo "Testing perf lock record and perf lock contention" - perf lock record -o ${perfdata} -- perf bench sched messaging > /dev/null 2>&1 + perf lock record -o ${perfdata} -- perf bench sched messaging \ + --group "$(nproc)" --thread "$(nproc)" > /dev/null 2>&1 # the output goes to the stderr and we expect only 1 output (-E 1) perf lock contention -i ${perfdata} -E 1 -q 2> ${result} if [ "$(cat "${result}" | wc -l)" != "1" ]; then --
Quoted 3 lines
2.48.1
Signature · 9 lines
Thomas Richter, Dept 3303, IBM s390 Linux Development, Boeblingen, Germany -- IBM Deutschland Research & Development GmbH Vorsitzender des Aufsichtsrats: Wolfgang Wendt Geschäftsführung: David Faller Sitz der Gesellschaft: Böblingen / Registergericht: Amtsgericht Stuttgart, HRB 243294
Namhyung Kim
maintainerHello,
Quoted 15 lines · Jan Polensky
On low-activity systems, the 'kernel lock contention analysis test' often fails due to insufficient system load. To address this, the test now increases load by using multiple groups and threads in 'perf bench sched messaging', scaled to the number of available CPUs.
Signed-off-by Jan Polensky <[email protected]>--- tools/perf/tests/shell/lock_contention.sh | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)tools/perf/tests/shell/lock_contention.sh +0
-0
tools/perf/tests/shell/lock_contention.sh check() { · line 48
Quoted 6 lines
tools/perf/tests/shell/lock_contention.sh +2
-1@@ -44,7 +44,8 @@ check() { test_record() { echo "Testing perf lock record and perf lock contention" - perf lock record -o ${perfdata} -- perf bench sched messaging > /dev/null 2>&1 + perf lock record -o ${perfdata} -- perf bench sched messaging \ + --group "$(nproc)" --thread "$(nproc)" > /dev/null 2>&1
$ perf bench sched messaging -h # Running 'sched/messaging' benchmark:
Usage: perf bench sched messaging <options>
-g, --group <n> Specify number of groups
-l, --nr_loops <n> Specify the number of loops to run (default: 100)
-p, --pipe Use pipe() instead of socketpair()
-t, --thread Be multi thread instead of multi processThe default value of --group option is 10 so it can be smaller on tiny systems. Note that it'd create 20 tasks for each group (I don't think we have an option to control this).
Also --thread option doesn't take an argument so the number will be ignored.
More importantly, there are several instances of this workload. Probably we want to change all of them by introducing a shell variable. But I'm a bit afraid how much overhead it'd bring up.
Thanks,
Namhyung
Quoted 6 lines
tools/perf/tests/shell/lock_contention.sh +0
-1# the output goes to the stderr and we expect only 1 output (-E 1) perf lock contention -i ${perfdata} -E 1 -q 2> ${result} if [ "$(cat "${result}" | wc -l)" != "1" ]; then --2.48.1
Hello Namhyung,
Quoted 9 lines · Namhyung Kim
Hello,
Quoted 7 lines · Jan Polensky
On low-activity systems, the 'kernel lock contention analysis test' often fails due to insufficient system load. To address this, the test now increases load by using multiple groups and threads in 'perf bench sched messaging', scaled to the number of available CPUs.
Signed-off-by Jan Polensky <[email protected]>---
[skip]
Quoted 3 lines
More importantly, there are several instances of this workload. Probably we want to change all of them by introducing a shell variable.
You're right — introducing a shell variable makes sense, especially since there are 13 identical calls. However, I’m slightly concerned that this might reduce readability. We might also consider an alternative solution to avoid this.
Quoted 1 line
But I'm a bit afraid how much overhead it'd bring up.
Indeed, using the --group parameter on large systems can lead to issues. For example:
[root@localhost perf]# ./perf stat -- perf bench sched messaging --group $(nproc) --thread # Running 'sched/messaging' benchmark:
perf: socketpair(): Too many open files
To avoid this, I suggest following the approach proposed by Thomas Richter ([email protected]), using the -p (pipe) mode instead:
perf bench sched messaging -p
The pipe-based solution generates sufficient load and lock contention for the test, while avoiding the use of sockets and their associated file descriptor limits.
I’ll incorporate these changes and send the next version shortly.
Thanks again for your feedback – it's much appreciated.
Best regards,
Jan
Namhyung Kim
maintainerQuoted 11 lines · Jan Polensky
Hello Namhyung,
Quoted 9 lines · Namhyung Kim
Hello,
Quoted 7 lines · Jan Polensky
On low-activity systems, the 'kernel lock contention analysis test' often fails due to insufficient system load. To address this, the test now increases load by using multiple groups and threads in 'perf bench sched messaging', scaled to the number of available CPUs.
Signed-off-by Jan Polensky <[email protected]>---
Quoted 8 lines
[skip]
Quoted 3 lines
More importantly, there are several instances of this workload. Probably we want to change all of them by introducing a shell variable.
You're right — introducing a shell variable makes sense, especially since there are 13 identical calls. However, I’m slightly concerned that this might reduce readability. We might also consider an alternative solution to avoid this.
We already have similar code in other places like record.sh.
Thanks,
Namhyung
Quoted 23 lines
Quoted 1 line
But I'm a bit afraid how much overhead it'd bring up.
Indeed, using the --group parameter on large systems can lead to issues. For example:
[root@localhost perf]# ./perf stat -- perf bench sched messaging --group $(nproc) --thread # Running 'sched/messaging' benchmark:
perf:socketpair(): Too many open filesTo avoid this, I suggest following the approach proposed by Thomas Richter ([email protected]), using the -p (pipe) mode instead:
perf bench sched messaging -p
The pipe-based solution generates sufficient load and lock contention for the test, while avoiding the use of sockets and their associated file descriptor limits.
I’ll incorporate these changes and send the next version shortly.
Thanks again for your feedback – it's much appreciated.
Best regards,
Jan