Messages in this thread | | | Date | Tue, 2 Oct 2012 18:30:02 +0200 | From | Andi Kleen <> | Subject | Re: [PATCHv2 00/14] perf diff: Factor diff command |
| |
> > Some minor issues I ran into (but no show stoppers): > > - The error messages for bad -c expressions could be better > > - I found the requirement for no space after -c unintuitive. > > I'll see to that > > > - It would be nice to have support for doing the bucketizing per line > > instead of per function. With a large function and/or inlining > > it's sometimes hard to identify the actual problem > > Acme recently added this for perf report. > > hm, I missed this one.. hopefully it should be no problem to add it
Here's more issues I found (again no show stopper):
For example you have a spinlock problem with different spinlock callers that you want to handle. Since the spinlocks are own functions which are mixed together cannot be easily ratiod.
This is probably a bit hard to handle, but one way would be to use a variant of the inclusive reporting Arun recently submitted. Define groups of callers that have a combined cost including all callees. Use this as the diff unit.
I really liked the oprofile behaviour of attributing lock costs to the caller by default. With that it would just work.
The other problem I ran into is that perf archive doesn't seem to work very well with kernels, so it's hard to move profiles from different kernels around to diff them (e.g. for a performance regression) One way around this would be options to diff to specify the vmlinux etc. manually
-Andi
| |