On 01/24/2011 07:42 PM, Deng-Cheng Zhu wrote:
Hi, David
This version does fix the problem with 'perf stat'. However, when working
with 'perf record', the following happened:
-sh-4.0# perf record -f -e cycles -e instructions -e branches \
-e branch-misses -e r12 find / -name "*sys*">/dev/null
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.001 MB perf.data (~53 samples) ]
I get the same thing. What happens if you supply either '-c xxx' or '-f
xxx'?
I get:octeon:~/linux/tools/perf# ./perf record -e cycles /bin/ls -l /
total 100
drwxr-xr-x 2 root root 4096 2010-11-12 11:39 bin
[...]
drwxr-xr-x 13 root root 4096 2007-05-25 12:28 var
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.002 MB perf.data (~82 samples) ]
Almost no samples as you got.
But if I do:
octeon:~/linux/tools/perf# ./perf record -F 100000 -e cycles /bin/ls -l /
total 100
drwxr-xr-x 2 root root 4096 2010-11-12 11:39 bin
[...]
drwxr-xr-x 13 root root 4096 2007-05-25 12:28 var
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.404 MB perf.data (~17653 samples) ]
Look many more samples!
The question is, what is it supposed to do?
If you can get a reasonable number of samples out if you supply -c or
-F, then I would argue that it is working and the default settings for
-F are not a good fit for your test case.
I have slightly changed the patch. You could try the attached version
instead and tell me the results.
David Daney
mips-perf-64bit.patch
Description: Text document
|