Re: [PATCH] improve oprofile on many-way systems

[Anton B. <an...@sa...>]

Hi,

> The patch below should help oprofile's execrable performance on machines
> with "lots" of CPUs somewhat, but it really needs testing to see how
> much better it does, if any. At least task exit should be considerably
> better now.

Sorry I havent got back to you. I finally got to give the first 2
patches a spin (attached and updated to apply to current 2.6). The
third patch ended up with a bunch of rejects and I didnt have time to
resolve it all.

I ran a fork/exit intensive benchmark (sdet) on a big machine.

No patch:
Enabling oprofile resulted in over a 10 times slow down in performance.

oprofile1.patch:
About 2% slowdown when oprofile enabled

oprofile2.patch:
About 5% slowdown when oprofile enabled

So I obviously like these patches :) Keep in mind sdet is pretty much a
worst case for fork/exec so Im not too upset about a 5% drop.

Anton