IO susbsystem profiling sampling rate

[chakenalli_nanjegowd]

Hello,

I am trying to measure utilization of the IO subsystem of a long running application with large number of threads on KNL. But data collection stops early due to the 500MB limit. I wanted to know is there a way I can increase the sampling rate, so that the application is profiled for the full duration.

Below is the command I am using and it looks like there is no knob for collect io to increase the sampling rate.

amplxe-cl -collect io -analyze-system -- application

Thank you,

Ram

[Vladimir_R_Intel]

Hi Ramachandra,

I suppose that increasing of sampling interval will not help you. But you can increase the data limit:

-data-limit=<integer> (500)   Limit the amount of raw data to be collected by

                              setting the maximum possible result size (in MB).

                              VTune Amplifier starts collecting data from the

                              beginning of the target execution and ends when

                              the limit for the result size is reached. For

                              unlimited data size, specify 0.

Also it will be helpful if you send us your result or at least the list of sizes of trace files (du -h -d 2 <your_result>/data.0/*)

BR,

Vladimir

[chakenalli_nanjegowd]

Hello Vladimir,

Thank you for the pointers. I used this work around and the data collection directory size turned out to be huge. (about 27GB). I am attaching the du log you requested.

Moreover in the amplxe-gui I don't see the disk input out histogram. (rest of the summary is fine). I just see a message "No data to show. The collected data is not sufficient" under disk IO histogram. The IO subsystem is a scratch directory mounted through NFS, so not sure if the VTUNE can capture it. Let me know if anything I missed in the command in the below command.

sudo amplxe-cl -data-limit=0 -collect io -analyze-system -- my_application -threads 16

Thank you,

Ram