Opened 3 years ago

Closed 3 years ago

#10247 closed task (worksforme)

5x increase in total memory allocation for `prof-doc-fib` test since 2011

Reported by: thomie Owned by:
Priority: high Milestone:
Component: Profiling Version: 7.10.1
Keywords: Cc:
Operating System: Unknown/Multiple Architecture: Unknown/Multiple
Type of failure: None/Unknown Test Case:
Blocked By: Blocking:
Related Tickets: Differential Rev(s):
Wiki Page:

Description

Commit c81e07063dd4f792d65f5933cfb906620d120b24 (#8811) makes a small change to the formatting of the profiling output. When updating the prof.sample files in the testsuite/tests/profiling/should_run directory, we noticed that memory allocation numbers for several tests have increased substantially since 2011.

-total alloc = 107,829,264 bytes  (excludes profiling overheads)
+total alloc = 538,951,032 bytes  (excludes profiling overheads)

The following tests need investigation:

  • T680.prof.sample (2x total alloc)
  • prof-doc-fib.prof.sample (5x)
  • prof-doc-last.prof.sample (3x)
  • scc003.prof.sample (2x)

Maybe it's nothing serious, but I'm setting priority to high since it seems like a regression.

Change History (3)

comment:1 Changed 3 years ago by rwbarton

I tried running the tests with 7.8.4 since I don't have HEAD built with profiling libraries at the moment.

I see almost exactly 2x factors on T680, prof-doc-last, and scc003. I also notice that these prof files are all from Oct 25-27 2011. Probably they were just generated on a 32-bit processor?

For prof-doc-fib I notice that the profiling output changed in a recent (not in 7.8.4) commit daed18c35cda114d8a5303bcb645195e1fd397e3, and the old file agrees with the above. That is, it showed almost exactly half the total allocation that I see, and it is from the same date range.

So there seems to be no alarm through 7.8.4. I'll try HEAD to see if there is any change.

comment:2 Changed 3 years ago by rwbarton

Now testing with HEAD, on a hunch I tried make fast TEST=prof-doc-fib (which only runs the profasm test way), and I got a number which (exactly!) matched the previous prof-doc-fib.prof.sample value. So I think that file was produced with profasm. (Would be nice to record this information...)

prof-doc-last really did regress from 7.8.4 to 7.10.1, though:

total alloc = 408,053,512 bytes  (excludes profiling overheads)  -- 7.8.4
total alloc = 648,053,536 bytes  (excludes profiling overheads)  -- 7.10.1

presumably due to b709f0a047dc036de15dc43d3b0ab88d3e32c5e6. But the extra allocations go away when optimizations are enabled, so I guess it's okay.

comment:3 Changed 3 years ago by rwbarton

Resolution: worksforme
Status: newclosed

I filed the last issue as #10260 so I'm going to close this--would really be nice to record more information about those prof.sample files though.

Note: See TracTickets for help on using tickets.