Vtune unable to analyze all captured data

[eduardoalberti]

Hello all,

I'm facing a possible issue during system analysis using Vtune 2024.2.1. When I try to run a Hotspot Analysis, or a System Overview, the Vtune isn't able to process all the data captured, creating an inaccurate visualization of the system. In example, when I run a hotspot analysis with 15 minutes duration time, the vtune only shows the CPU activity for the first 100 seconds. Sometimes even less then 100s are showed on the graphs. The data table is also affected.

I ran vtune agent as a standalone system, at localhost, I mean without run the server and using the command
 ./vtune -collect hotspots -knob sampling-mode=hw -knob enable-stack-collection=true -knob sampling-interval=5 -knob stack-size=2048 -knob enable-characterization-insights=true -duration=900 -user-data-dir /tmp/intel

I tried also to run remotely, using the Vtune GUI. The results are the same.

[yuzhang3_intel]

Are you sure your command line can work? call stack collection is not available when system-wide profiling is enabled.

vtune -collect hotspots -knob sampling-mode=hw -knob enable-stack-collection=true -d 10

vtune: Warning: Stack flow analysis on this platform is limited to the hardware LBR-based stack type that has a depth limitation.

vtune: Error: Call stack collection is not available when system-wide profiling is enabled. Disable Collect stacks option in HOW pane. Alternatively, disable system-wide profiling in WHAT pane. In Launch Application mode, uncheck Analyze system-wide option under advanced settings. You cannot perform call stack collection in Profile System mode.

[eduardoalberti]

Hi, 
Thank you for the answer.

Yes, I'm sure that the command line is correct. I changed only the duration to show you the command

The GUI also allow the usage of the stack collector

[yuzhang3_intel]

I verified locally using the below command line to profile one application for 900s, you can see it works fine. 

yuzhang3@yuzhang3-10710:~/workspace/test_dir/loop_src$ vtune -collect hotspots -knob sampling-mode=hw -knob enable-stack-collection=true -knob sampling-interval=5 -knob stack-size=2048 -knob enable-characterization-insights=true -duration=900 ./test_loop

Profiling system-level for 900s w/o call stack collection can work fine as well.

yuzhang3@yuzhang3-10710:~/workspace/test_dir/loop_src$ vtune -collect hotspots -knob sampling-mode=hw -knob enable-characterization-insights=true -knob sampling-interval=5 -duration=900

[eduardoalberti]

Sorry to insist in the same topic, but I tried to run the test more times, with and without stack collection. 
When I run the collection without the stack the summary seems to be more realistic on the CPU time:

But when I change to the "Bottom-up" CPU visualization I see the same problem, the CPU graph doesn't show the entire activity.

I tried to run the test with different durations, and I noted that if the test has more than 3 minutes the problem appears.

[yuzhang3_intel]

Could you please attach the VTune data? Let me check, thanks.

[Jennifer_D_Intel]

Can you also try adding this knob:

-finalization-mode=full

By default, VTune does fast finalization, which could result in missing data if there is a very large number of samples. Seeing as how your collection shows almost 120,000 threads, this seems possible.