Skip to contentSkip to conversation

perf test: Adjust test case perf record offcpu profiling tests for s390

PATCH Applied Thomas Richter 3 participants 3 messages Nov 6, 2023 – Nov 8, 2023 perf events

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 perf test: Adjust test case perf record offcpu profiling tests for s390

On s390 using linux-next the test case
87: perf record offcpu profiling tests
fails. The root cause is this command

 # ./perf  record --off-cpu -e dummy -- ./perf bench sched messaging -l 10
 # Running 'sched/messaging' benchmark:
 # 20 sender and receiver processes per group
 # 10 groups == 400 processes run
     Total time: 0.231 [sec]

[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.077 MB perf.data (401 samples) ] #

It does not generate 800+ sample entries, on s390 usually around 40[1-9], sometimes a few more, but never more than 450. The higher the number of CPUs the lower the number of samples.

Looking at function chain
bench_sched_messaging()
+--> group()
the senders and receiver threads are created. The senders and receivers call function ready() which writes one bytes and wait for a reply using poll system() call.

As context switches are counted, the function ready() will trigger a context switch when no input data is available after the write system call. The write system call does not trigger context switches when the data size is small. And writing 1000 bytes (10 iterations with 100 bytes) is not much and certainly won't block.

The 400+ context switch on s390 occur when the some receiver/sender threads call ready() and wait for the response from function bench_sched_messaging() being kicked off.

Lower the number of expected context switches to 400 to succeed on s390.

---
 tools/perf/tests/shell/record_offcpu.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/tests/shell/record_offcpu.sh +2 -2
index a1ef8f0d2b5c..67c925f3a15a 100755
@@ -77,9 +77,9 @@ test_offcpu_child() {
     err=1
     return
   fi
-  # each process waits for read and write, so it should be more than 800 events
+  # each process waits at least for poll, so it should be more than 400 events
   if ! perf report -i ${perfdata} -s comm -q -n -t ';' --percent-limit=90 | \
-    awk -F ";" '{ if (NF > 3 && int($3) < 800) exit 1; }'
+    awk -F ";" '{ if (NF > 3 && int($3) < 400) exit 1; }'
   then
     echo "Child task off-cpu test [Failed invalid output]"
     err=1
Signature · 1 line
2.41.0
maintainer Acked-by on v2 1/1
#1 lore raw

Hello,

Quoted 36 lines · Thomas Richter

On s390 using linux-next the test case
87: perf record offcpu profiling tests
fails. The root cause is this command

 # ./perf  record --off-cpu -e dummy -- ./perf bench sched messaging -l 10
 # Running 'sched/messaging' benchmark:
 # 20 sender and receiver processes per group
 # 10 groups == 400 processes run
     Total time: 0.231 [sec]

[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.077 MB perf.data (401 samples) ] #

It does not generate 800+ sample entries, on s390 usually around 40[1-9], sometimes a few more, but never more than 450. The higher the number of CPUs the lower the number of samples.

Looking at function chain
bench_sched_messaging()
+--> group()
the senders and receiver threads are created. The senders and receivers call function ready() which writes one bytes and wait for a reply using poll system() call.

As context switches are counted, the function ready() will trigger a context switch when no input data is available after the write system call. The write system call does not trigger context switches when the data size is small. And writing 1000 bytes (10 iterations with 100 bytes) is not much and certainly won't block.

The 400+ context switch on s390 occur when the some receiver/sender threads call ready() and wait for the response from function bench_sched_messaging() being kicked off.

Quoted 7 lines

Lower the number of expected context switches to 400 to succeed on s390.

Thanks,
Namhyung

Quoted 23 lines
---
 tools/perf/tests/shell/record_offcpu.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/tests/shell/record_offcpu.sh +2 -3
index a1ef8f0d2b5c..67c925f3a15a 100755
@@ -77,9 +77,9 @@ test_offcpu_child() {
     err=1
     return
   fi
-  # each process waits for read and write, so it should be more than 800 events
+  # each process waits at least for poll, so it should be more than 400 events
   if ! perf report -i ${perfdata} -s comm -q -n -t ';' --percent-limit=90 | \
-    awk -F ";" '{ if (NF > 3 && int($3) < 800) exit 1; }'
+    awk -F ";" '{ if (NF > 3 && int($3) < 400) exit 1; }'
   then
     echo "Child task off-cpu test [Failed invalid output]"
     err=1
--

2.41.0

maintainer Applied
#2 lore raw

Em Mon, Nov 06, 2023 at 07:21:35AM -0800, Namhyung Kim escreveu:

Quoted 10 lines · Thomas Richter
Quoted 7 lines

Lower the number of expected context switches to 400 to succeed on s390.

Thanks, applied to perf-tools-next.

- Arnaldo