Re: [PATCH] [15/58] i386: Rewrite sched_clock (cmpxchg8b)

[Mathieu Desnoyers <compudj@krystal.dyndns.org>]

* Nick Piggin (nickpiggin@yahoo.com.au) wrote:
> Mathieu Desnoyers wrote:
> 
> >I tried it with and without the LOCK prefix on my Pentium 4.
> >
> >Locked cmpxchg8b : 90 cycles
> >Non locked cmpxchg8b: 30 cycles
> >sti: 166 cycles
> >cli: 159 cycles
> >
> >So, hrm, even if we use the locked version, it is still much faster than
> >the sti/cli. I am thoughtful about the comment in asm-i386/system.h:
> 
> Curious: what does it look like if the memory is not in cache? I
> found that cmpxchg is relatively slower than other rmw instructions
> in that case.
> 

Actually, I have just seen that cmpxchg64 and cmpxchg64_local are
doing exactly this and they are already implemented in asm-i386/system.h.

A quick test: I am doing clflush in a loop (substracting its time from the
following loops) to have a memory hit when I do cmpxchg. This is the
result of just the cmpxchg8b:

non locked cmpxchg8b: 583.37 cycles
locked cmpxchg8b: 650.48 cycles
rmw in 3 operations: 581.43 cycles

So the locked cmpxchg is 67 cycles slower than the non locked cmpxchg,
which fits with my 30 vs 90 cycles. rmw is a tiny bit faster than
cmpxchg8b (2 cycles), but nothing to call home about.

Mathieu

-- 
Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/