Re: Another Performance Regression in write() syscall

[Ingo Molnar <mingo@elte.hu>]

* Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Tue, 24 Feb 2009, Dave Hansen wrote:
> > 
> > Yeah, that's a good point.  Are we sure that's what is 
> > happening here, though?  That's one thing a profile would 
> > hopefully help with.
> 
> One thing to note is that _if_ it's purely an issue of 
> nontemporal stores vs normal stores, then profiling is very 
> likely going to be almost entirely useless. You'll get 
> "results", but the results have nothing what-so-ever to do 
> with reality or anything interesting.
> 
> The nontemporal stores may stand out in the profiles, but the 
> actual performance impact will be all about whether totally 
> unrelated code got cache misses or not. Quite often those 
> cache misses will also be in user mode, and very possibly in 
> other processes.
> 
> So profiles can certainly be interesting, but if Salman says 
> that his patch makes a difference for his benchmark, then 
> profiling is almost certainly not interesting FOR THAT PATCH. 
> It's interesting mainly as a way to look at whether there are 
> then also _other_ issues that are worth addressing (ie the 
> whole atime thing is in a whole different dimension and an 
> independent issue).

a 'perfstat' run would certainly be interesting (for cases where 
a pure /usr/bin/time run is inconclusive), comparing the 
unpatched and patched kernel.

That way we can see summary counts for the whole workload, like:

 -----------------------------------------------
 | Performance counter stats for './mmap-perf' |
 -----------------------------------------------
 |                |
 |  x86-defconfig |   PARAVIRT=y
 |------------------------------------------------------------------
 |
 |    1311.554526 |  1360.624932  task clock ticks (msecs)    +3.74%
 |                |
 |              1 |            1  CPU migrations
 |             91 |           79  context switches
 |          55945 |        55943  pagefaults
 |    ............................................
 |     3781392474 |   3918777174  CPU cycles                  +3.63%
 |     1957153827 |   2161280486  instructions               +10.43%
 |       50234816 |     51303520  cache references            +2.12%
 |        5428258 |      5583728  cache misses                +2.86%
 |
 |      437983499 |    478967061  branches                    +9.36%
 |       32486067 |     32336874  branch-misses               -0.46%
 |                |
 |    1314.782469 |  1363.694447  time elapsed (msecs)        +3.72%
 |                |
 -----------------------------------

Such a comparison of would certainly be more meaningful for such 
things than a profile.

Salman, if you are interested in doing a perfstat comparison, 
just pick up a tip:master kernel [perfcounters are 
default-enabled in it]:

   http://people.redhat.com/mingo/tip.git/README

and run perfstat on it (as root, to get the kernel-mode counts 
too):

   http://redhat.com/~mingo/perfcounters/perfstat.c

	Ingo