vtuen problem：why 'input and output' and macc analyzes did not capture all processes

[YitianZheng]

If someone could provide assistance, I would be deeply grateful: 

Background: I have written a script named fillrandom_no_group.sh which connects 16 rocksdb benchmark instances, ./db_bench, with an '&' between them. By calling fillrandom_no_group.sh, 16 db_bench processes are sequentially launched, which will subsequently run concurrently.

The server configuration is as follows: Ubuntu 22.04 CPU: INTEL(R) XEON(R) GOLD 6530 (Logical CPU count: 128).

My problem:

When running fillrandom_no_group.sh, I observed in htop that all 16 processes were running, each containing multiple threads. When I ran 'hotspots' analysis, the output from VTune largely matched what I observed in htop (see Figure 1).

 

However, when I ran 'input and output' analysis, I noticed significant discrepancies in CPU utilization compared to 'hotspots' analysis (see Figure 2). This was reflected in CPU time, total thread count, and utilization.

Moreover, in the bottom-up dropdown menu, only 5 db_bench processes were visible (see Figure 3). It seems that the 'input and output' analysis did not capture all processes invoked by the analyzed script, resulting in inaccuracies in CPU-related data. What could be the reason for this? How can I adjust for correct testing? 

[yuzhang3_intel]

Could you attach IO VTune data and Hotspots VTune data as well?

[YitianZheng]

Sorry, are you referring to the file ending in. vtune, the entire folder, or something else?

[yuzhang3_intel]

After profiling, you can get some VTune data folders, like r00hs, r01io, etc. The folders contain some information.