Re: lmbench regression due to cond_resched nullification change 26-rc5 vs. 25

[Christian Borntraeger]

Am Freitag, 20. Juni 2008 schrieben Sie:
> But quite frankly, regardless of exactly why it happens, it absolutely 
> makes no sense to even have that cond_resched() in the _innermost_ loop - 
> the one that is called for every single fd. It's much better to move the 
> conditional reschedule out a bit.
> 
> That inner loop was very much designed to compile into nice assembly 
> language, and it's possible that the cond_resched() simply causes extra 
> register pressure and keeps us from keeping the bitmasks in registers etc.
> 
> So this trivial patch, which moves the cond_resched() to be outside the 
> single-word select() loop, would be a good idea _regardless_. Does it make 
> any difference for you? If it's purely a "code generation got worse" 
> issue, it should help a bit.

Yes, it helps. I am now at 24 microseconds which is very close to the original 
time for lat_select.

Thanks

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/