I'm having some problems with vtune on Intel DevCloud.
What I want to do is performing analysis from command line in devcloud and then download results on local machine.
When I run the gpu-offload analysis and open results with tune gui, the global working size of my kernel visualised doesn't match with range<> passed to my kernel.
The other problem is that if I run the source analysis, the application exits as soon as it reaches the gpu kernel.
The command used in this case is:
vtune -collect gpu-hotspots -knob profiling-mode=source-analysis ./myapp
Thanks for reaching out to us. Could you please provide the following information with us, so we can debug the issue in detail.
Case 1: GPU Offload-global working size of kernel issue
1. Please check the version of vtune that you are using. If it's oldest one, please upgrade to latest beta version of Vtune.
2. Are you using same version of vtune on your local machine as that of vtune installed in Devcloud?
3. If possible, please provide the source code or workload with us to investigate more on the issue.
As a workaround, instead of downloading the result directory on to your local machine you can use Intel® VTune™ Profiler in a web server mode. This will launch vtune profiler in browser on your host machine. Please refer to the below documentation for more details on launching the web server & running analysis. Link: https://software.intel.com/content/www/us/en/develop/documentation/vtune-help/top/launch/web-server-...
Case 2: Source analysis problems
We tried out the command you shared for the matrix application in GPU nodes s001-n191 & s001-n192. It seems to be working without any issues for us. So, the issue might be with the node you are running. Could please try it in the above mentioned nodes. Please let us know the updates.
Glad to know that your issue is resolved.
For your question, Vtune web server work on Devlcoud: Yes, definitely it will work from Devcloud. You should launch Vtune web server from Devcloud by running the vtune-backend command. Please look into the below steps to launch the server.
VTune Profiler Server outputs a URL to access the GUI. You can open the URL in browser on your local machine. In Devcloud, you may need to do two-way tunneling to open the URL. For two-way tunneling , open a new prompt and tunnel to login node of the Devcloud using the below command:
ssh -L 8880:localhost:8880 devcloud_proxy
Then, tunnel to compute node as
ssh -L 8880:localhost:8880 s001-n058
devcloud_proxy :- Host name
8880 :- Port in which vtune web sever is running
s001-n058:- compute node in which your vtune web server has started
Please follow the further steps in the documentation shared.
Since your initial query is resolved, we won't be monitoring this thread anymore. Kindly raise a new thread if you need further assistance.