Analyzers
Talk to fellow users of Intel Analyzer tools (Intel VTune™ Profiler, Intel Advisor)
5250 Diskussionen

Vtune multiple-run crashing

Roshantha_M_
Einsteiger
1.142Aufrufe

Hi,

When I try to run Vtune via CLI and enable allow-multiple-runs, the execution crashes. Works fine on single runs.

command :

root@laptop:/opt/intel/vtune_amplifier_xe# /opt/intel/vtune_amplifier_xe_2016.2.0.444464/bin64/amplxe-cl -collect memory-access -knob analyze-mem-objects=true -target-duration-type=veryshort -allow-multiple-runs -data-limit=501 -app-working-dir /home/rosh/Documents/openHEVC/build -- /home/rosh/Documents/openHEVC/build/hevc -i ../../ffmpeg_work/x265_encoded_vids/video.x265 -f 1 -p 1 -n

The execution duration of the app im profiling is less than 10 seconds, hence I've chosen 'veryshort'

Same behaviour when using the GUI.

running Linux (Ubuntu 14.04).

I've attached the crash log. What could be the problem ?

Many thanks

0 Kudos
3 Antworten
Dmitry_P_Intel1
Mitarbeiter
1.142Aufrufe

Hello,

We could reproduce the issue with multiple run, it is connected with instrumentation of memory objects through -knob analyze-mem-objects=true in this mode, will keep you posted.

Thanks & Regards, Dmitry

Roshantha_M_
Einsteiger
1.142Aufrufe

Thank you Dmitry.

Okay, thats good. I was worried, it was something wrong with my system settings/application.

As a workaround until this is fixed, can I manually run the same command multiple times and get the average for each reported metric ? is that how the multiple run calculates the results ? as averages ? how many runs does a multiple-run do ?

thanks

TimP
Geehrter Beitragender III
1.142Aufrufe

If you don't "allow multiple runs," you are in the "multiplexed" mode where VTune shifts among groups of counters, assuming that you have specified more than it can use simultaneously.  If you were not using "analyze mem objects," the multiple runs mode would likely be best for short runs such as you have, even if you have set the shorter sampling interval.  You could specify few enough events per case to avoid the multiplexing, if you are willing to undertake combining multiple runs yourself.

Antworten