modifty sleepers to have easier evaluation of interaction between cpu freq scaling

and scheduler lock (pipe based or futex based)

See http://www.valgrind.org/docs/manual/manual-core.html#manual-core.pthreads_perf_sched
for background info about cpu freq scaling and valgrind thread scheduler.

To reproduce the interaction, do:
for sched in --fair-sched=yes --fair-sched=no
do
  for affinity in 0 1
  do
    echo $sched $affinity
    time ./vg-in-place $sched -q ./gdbserver_tests/sleepers 1000000 0 1000 B-B-B-B- $affinity
  done
done

which gives the below output (intel core i5-6402P, debian 8, kernel  3.16.0).
In summary: the fair scheduler is fair, the pipe based scheduler
can be really unfair (e.g. with --fair-sched=no and no affinity,
2 threads are finishing their work, while the 2 other threads are
starting their work only after the first 2 have fully finished).
The difference in timing is significant : 1m14s versus around 47 seconds.

  Note: If the governor is set to performance, strangely, the time needed for
  --fair-sched=no increases slighltly (to around 48 seconds).
  The time for --fair-sched=yes with or without affinity is then also
  to around 48 seconds.

Below is timing with on-demand governor:

--fair-sched=yes 0
loops/sleep_ms/burn/threads_spec/affinity:  1000000 0 1000 B-B-B-B- 0
Brussels ready to sleep and/or burn
London ready to sleep and/or burn
Petaouchnok ready to sleep and/or burn
main ready to sleep and/or burn
Brussels finished to sleep and/or burn
London finished to sleep and/or burn
Petaouchnok finished to sleep and/or burn
main finished to sleep and/or burn

real	1m14.582s
user	1m14.348s
sys	0m0.204s
--fair-sched=yes 1
loops/sleep_ms/burn/threads_spec/affinity:  1000000 0 1000 B-B-B-B- 1
Brussels ready to sleep and/or burn
London ready to sleep and/or burn
Petaouchnok ready to sleep and/or burn
main ready to sleep and/or burn
main finished to sleep and/or burn
Brussels finished to sleep and/or burn
Petaouchnok finished to sleep and/or burn
London finished to sleep and/or burn

real	0m46.785s
user	0m46.756s
sys	0m0.032s
--fair-sched=no 0
loops/sleep_ms/burn/threads_spec/affinity:  1000000 0 1000 B-B-B-B- 0
Brussels ready to sleep and/or burn
Brussels finished to sleep and/or burn
London ready to sleep and/or burn
London finished to sleep and/or burn
Petaouchnok ready to sleep and/or burn
main ready to sleep and/or burn
Petaouchnok finished to sleep and/or burn
main finished to sleep and/or burn

real	0m47.742s
user	0m48.224s
sys	0m0.084s
--fair-sched=no 1
loops/sleep_ms/burn/threads_spec/affinity:  1000000 0 1000 B-B-B-B- 1
Brussels ready to sleep and/or burn
London ready to sleep and/or burn
Petaouchnok ready to sleep and/or burn
main ready to sleep and/or burn
Brussels finished to sleep and/or burn
London finished to sleep and/or burn
main finished to sleep and/or burn
Petaouchnok finished to sleep and/or burn

real	0m46.601s
user	0m46.568s
sys	0m0.036s



git-svn-id: svn://svn.valgrind.org/valgrind/trunk@16251
This commit is contained in:
Philippe Waroquiers 2017-02-19 11:23:46 +00:00
parent 91e7677668
commit d7ccda99e6
11 changed files with 13 additions and 13 deletions

View File

@ -1,4 +1,4 @@
loops/sleep_ms/burn/threads_spec: 1 0 2000000000 ------B-
loops/sleep_ms/burn/threads_spec/affinity: 1 0 2000000000 ------B- 1
main ready to sleep and/or burn
pid .... Thread .... inferior call pushed from gdb in mcinfcallRU.stdinB.gdb
Reset valgrind output to log (orderly_finish)

View File

@ -2,7 +2,7 @@
prog: sleepers
# We would like to use B-B-B-B- instead of ------B- but this gives
# too much dependencies to the scheduler fairness.
args: 1 0 2000000000 ------B-
args: 1 0 2000000000 ------B- 1
vgopts: --tool=memcheck --vgdb=yes --vgdb-error=0 --vgdb-prefix=./vgdb-prefix-mcinfcallRU
# filter_gdb to replace pid and Thread numbers in the output of the program:
stderr_filter: filter_gdb

View File

@ -1,4 +1,4 @@
loops/sleep_ms/burn/threads_spec: 100 100000000 1000000000 -S-S-SB-
loops/sleep_ms/burn/threads_spec/affinity: 100 100000000 1000000000 -S-S-SB- 1
Brussels ready to sleep and/or burn
London ready to sleep and/or burn
Petaouchnok ready to sleep and/or burn

View File

@ -3,7 +3,7 @@
prog: sleepers
# We would like to have two threads running (i.e. -S-SB-B-)
# but this introduces too much dependencies to scheduler fairness.
args: 100 100000000 1000000000 -S-S-SB-
args: 100 100000000 1000000000 -S-S-SB- 1
vgopts: --tool=memcheck --vgdb=yes --vgdb-error=0 --vgdb-prefix=./vgdb-prefix-mcinfcallWSRU
# We need a non buggy gdb.step on arm thumb.
# Disable on Darwin: inferior call rejected as it cannot find malloc.

View File

@ -1,7 +1,7 @@
# test that vgdb can invoke a process when all threads are in Runnable or Yielding mode
# If the test goes wrong, it might consume CPU during a long time.
prog: sleepers
args: 1 0 1000000000 B-B-B-B-
args: 1 0 1000000000 B-B-B-B- 1
vgopts: --tool=memcheck --vgdb=yes --vgdb-prefix=./vgdb-prefix-mcinvokeRU
stderr_filter: filter_make_empty
# as the Valgrind process is always busy, we do not need the vgdb.invoker prereq.

View File

@ -3,7 +3,7 @@ Nulgrind, the minimal Valgrind tool
(action at startup) vgdb me ...
loops/sleep_ms/burn/threads_spec: 1000000000 1000000000 1000000000 BSBSBSBS
loops/sleep_ms/burn/threads_spec/affinity: 1000000000 1000000000 1000000000 BSBSBSBS 1
Brussels ready to sleep and/or burn
London ready to sleep and/or burn
Petaouchnok ready to sleep and/or burn

View File

@ -9,7 +9,7 @@
# This test is disabled on Solaris because modifying select/poll/ppoll timeout
# has no effect if a thread is already blocked in that syscall.
prog: sleepers
args: 1000000000 1000000000 1000000000 BSBSBSBS
args: 1000000000 1000000000 1000000000 BSBSBSBS 1
vgopts: --tool=none --vgdb=yes --vgdb-error=0 --vgdb-prefix=./vgdb-prefix-nlcontrolc
stderr_filter: filter_stderr
# Bug 338633 nlcontrol hangs on arm64 currently.

View File

@ -4,7 +4,7 @@
# But if this signal is masked, then vgdb does not recuperate the control
# and Valgrind dies. See function give_control_back_to_vgdb in m_gdbserver.c
prog: sleepers
args: 1 10000000 0 -S-S-S-S
args: 1 10000000 0 -S-S-S-S 1
vgopts: --tool=none --vgdb=yes --vgdb-error=0 --vgdb-prefix=./vgdb-prefix-nlsigvgdb
stderr_filter: filter_stderr
prereq: test -e gdb -a -f vgdb.invoker

View File

@ -3,7 +3,7 @@ Nulgrind, the minimal Valgrind tool
(action at startup) vgdb me ...
loops/sleep_ms/burn/threads_spec: 1000000000 1000000000 1 BSBSBSBS
loops/sleep_ms/burn/threads_spec/affinity: 1000000000 1000000000 1 BSBSBSBS 1
Brussels ready to sleep and/or burn
London ready to sleep and/or burn
Petaouchnok ready to sleep and/or burn

View File

@ -3,7 +3,7 @@
# vgdb must queue these signals and deliver them before PTRACE_DETACHing.
# sleepers is started with argument so that it will mostly sleep.
prog: sleepers
args: 1000000000 1000000000 1 BSBSBSBS
args: 1000000000 1000000000 1 BSBSBSBS 1
vgopts: --tool=none --vgdb=yes --vgdb-error=0 --vgdb-prefix=./vgdb-prefix-nlvgdbsigqueue
stderr_filter: filter_stderr
prereq: test -e gdb -a -f vgdb.invoker

View File

@ -137,7 +137,7 @@ int main (int argc, char *argv[])
pthread_t ebbr, egll, zzzz;
struct spec b, l, p, m;
char *some_mem __attribute__((unused)) = malloc(100);
setaffinity();
if (argc > 5 && atoi(argv[5])) setaffinity();
setup_sigusr_handler();
if (argc > 1)
loops = atoi(argv[1]);
@ -153,8 +153,8 @@ int main (int argc, char *argv[])
else
threads_spec = "BSBSBSBS";
fprintf(stderr, "loops/sleep_ms/burn/threads_spec: %d %d %d %s\n",
loops, sleepms, burn, threads_spec);
fprintf(stderr, "loops/sleep_ms/burn/threads_spec/affinity: %d %d %d %s %d\n",
loops, sleepms, burn, threads_spec, argc > 5 && atoi(argv[5]));
fflush(stderr);
b.name = "Brussels";