VTune hangs Bottom-up

[jimdempseyatthecove]

I am doing VTune Basic Hotspot runs on 256 threads (KNL), 180 seconds runtime. (V17.0u4, on Linux)

When I finally get to the Summary page (long time on KNL), I can then either click on one of the different tabs, or click on a highlighted link of the handful of most active routines. When I click on Bottom-up tab, this often never completes. If instead I click on a highlighted link of the handful of most active routines it often succeeds. (although it is hanging as I type this). The Bottom-up window displays the Function call stack and information, but is disabled (greyed out). IOW I cannot brows the Function/Call stack.

If during Hang, I click on Top-down Tree tab, it will break out of the hang, and display the tree. Browsing via top-down is less convenient than using the bottom-up.

Jim Dempsey

[Vitaly_S_Intel]

Hi Jim,

Thanks for reporting this issue and sorry that you're unable to use the tool.

If you could share result directory with us, this would allow us to triage it quickly and hopefully fix in next releases.

If this is not possible, can you please run the following command after 1 minute of reproduced hang of Bottom-Up? I mean, keep VTune running, and in parallel run the following command from the terminal in the directory you're allowed to write files:

> <VTune install dir>/bin64/amplxe-feedback.exe -create-bug-report hang

It'll collect system info (you can disable this via adding '-no-system-info' option) and capture VTune process dump. As result you'll get hang.zip in the current directory. There will be only text files inside this archive - you can explore their content and remove everything you treat as sensitive data.

Please send hang.zip file to me via private message.

BTW, are you able to use Advanced Hotspots?

Thanks!

[jimdempseyatthecove]

>>BTW, are you able to use Advanced Hotspots?

On KNL (CentOS), with OpenMP running 256 threads, and SSH'ed from Windows 7, and using Xming to display on 4K display on W7 system, Basic Hotspots takes exceptionally long to generate the results database. Using Advanced Hotspots makes this about 10x longer. IOW Advanced Hotspots is not usable excepting for when I really, really need the information. I haven't done Advanced Hotspots recently, but it too had Bottom-up issues, Top-down worked. I will try again when the report finishes.

I am building the hang file now. (this takes a long time too, over 10 minutes and still running).

I would appreciate it if your VTune development team would increase and improve threading within VTune. While it is building the database, the system monitor shows one thread totally active, and maybe 2 or 3 other threads intermittent (I have several applications open but not running).

I suppose I could buy a 16-core Core i9-7960X Skylake with AVX-512 so I can "sort of" simulate the KNL environment at faster speed (and lower core/thread count). But this is an additional expense.

Jim Dempsey

[PAVEL_G_Intel]

Hi Jim,

Knights has pure single thread performance but VTune Amplifier finalization step is commonly single thread. It caused by dependencies in the collected data. So I don't recommend to run the data processing on KNL.

There are two ways to solve your problem:
1) Collect data on KNL and open result on another machine. Intel® Xeon Phi™ Processor (Code Name: Knights Landing) Targets
2) Setup host-target collection. Remote Linux* Target Setup 

I think that the second way is easier. What way do you prefer?

[Vitaly_S_Intel]

Hi Jim,

I confirm I received amplxe-feedback generated file from you and it contains all the data. We're analyzing it already.

I agree with Pavel's suggestion on trying to run VTune on your host system where VTune is supposed to run much faster.

Can you also please share what is the size of result directory (basic hotspots and advanced hotspots) which is causing these significant delays on loading hotspots view?

[jimdempseyatthecove]

Attached is .zip of

      ls -1 -l -R engine

Jim Dempsey

[Vitaly_S_Intel]

Hi Jim,

I don't see an attachment. I think you could just share approximate size of result directory..

[jimdempseyatthecove]

I placed it in the protected forum attachment (Sable Falls Users) area. Do you need it in the public forum?

Jim Dempsey