[ oprofile-Bugs-1561895 ] core_2 unit_mask/events nits

[SourceForge.net <no...@so...>]

Bugs item #1561895, was opened at 2006-09-20 04:02
Message generated for change (Comment added) made by movement
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=116191&aid=1561895&group_id=16191

Please note that this message will contain a full copy of the comment thread,
including the initial issue submission, for this request,
not just the latest update.
Category: None
Group: None
Status: Open
>Resolution: Fixed
Priority: 5
Submitted By: Dean Gaudet (dgaudet)
Assigned to: Nobody/Anonymous (nobody)
Summary: core_2 unit_mask/events nits

Initial Comment:
this is based off the 0.9.2.1rc1 core_2 events file...

the default mask for L2_RQSTS is 0xcf -- this double
counts references because oprofile is generally running
on all cores.  the default should probably be 0x4f. 
i've verified this with a multithreaded microbenchmark
issuing random walks on all processors and the results
for 0x4f are more in-line with what i'd expect.

i've attached a patch to correct this and several other
unit_mask problems.

-dean


----------------------------------------------------------------------

>Comment By: John Levon (movement)
Date: 2006-09-21 16:12

Message:
Logged In: YES 
user_id=53034

I've applied both after bcrl reviewed them.

The multiple events with the same number must be fixed by
adding synthetic event numbers and decoding them in-kernel.

----------------------------------------------------------------------

Comment By: Dean Gaudet (dgaudet)
Date: 2006-09-21 00:09

Message:
Logged In: YES 
user_id=408455

in 0.9.2.1rc1 there are two L2_RQSTS events listed for
core_2.  one is with um:core_mesi, the other with
um:core_prefetch_mesi... unfortunately since the prefetch
one comes second you can't specify prefetch masks for L2_RQSTS.

the patch i've attached also renames the two
L2_RQSTS.SELF.etcetc events to match what the ISA calls the
two events -- LLC_REFS and LLC_MISSES... this seems like it
would be preferred on the assumption they may some day add
an L3 cache and the architectural "LLC" events would
probably then refer to L3.

unfortunately it seems opreport only looks at the first 0x2e
event when decoding things so it doesn't display the handy
LLC_REFS/LLC_MISSES monikers.

anyhow, either this patch or one supporting
core_prefetch_mesi for L2_RQSTS would be nice...

thanks
-dean

----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=116191&aid=1561895&group_id=16191