inspector out of disk space on /home

inspector out of disk space on /home

Tim Prince的头像

Inspector is causing linux (rhel6.2 x86_64) to report out of disk space on /home and stopping with internal error.

I don't see any explicit indication of where it is writing data.  Only /home reports full but I don't see where it may have written data. After inspector quits /home reports only 3% full.   For good luck, I'm moving /opt/intel over to /home (2nd disk) and symlinking it back to the standard installation directory,

According to the comments I get on the screen, I need the full location of data races including stack, which causes this failure fairly quickly.  I can reduce aggressiveness and run over 25 minutes but still hit the same failure.

4 帖子 / 0 new
Tim Prince的头像

Oops, I meant out of data space on /.  Moving /opt/intel to the other disk seems to have helped, but it seemed like a wild guess.

Peter Wang (Intel)的头像

As my experience if you might try below:

1. Use options "-user-data-dir", "-result-dir" which partition has big size (RN said, at least 4GB)

2. Use option "-knob stack-depth=1" to reduce overhead, and result directory size will be small.

3. Don't check whole program, especially target application is huge. Use "Select analysis start location with debugger" - it means that analysis is disabled and execution (default "gdb") stops at application's entry, then you can set breakpoint so when execute to this bp, you can start analysis. In this mode, it is flexible for the user to analyze interest of code.

Tim Prince的头像

It seems threading errors which weren't diagnosed by Inspector contributed to unsatisfactory behavior.

It's not clear why red hat linux was reporting out of space on / when Inspector should not have been writing files there, and the hinted at requirement to have 4GB free disk space was observed.

Inspector with "normal" setting reports threads corrupting each others' stacks but at "thorough" setting this report doesn't come out, and Inspector reports setting itself back to normal later in order to continue the run.

Several races were specifically diagnosed by Inspector and could be avoided and get the application running on host, although not as fast as pure MPI.  The hope that this would allow working on MIC with MPI+OpenMP so far hasn't come true.