Re: aio is unlikely

[Nick Piggin <nickpiggin@yahoo.com.au>]

Jeff Garzik wrote:
> Andrew Morton wrote:
>
> > On Fri, 18 May 2007 17:54:32 -0400
> > Phillip Susi <psusi@cfl.rr.com> wrote:

> > > But as Jeff said, that's not what unlikely is for.  It should only be 
> > > used when it is unlikely for everybody, all the time, because when it 
> > > is right, it helps rather little, but when it is wrong, it hurts a lot.
> >
> >
> > It does?  Tell us more.
>
>
> It is difficult to quantify either way.  The details are both 
> CPU-specific and compiler-specific.  The best information can be culled 
> from the gcc list archives, which is where I obtained my knowledge on 
> the subject (which is now ~2 years old).
>
> Under the hood, likely() and unlikely() are implemented as percentage 
> predictions.  likely() is implemented in the kernel as a 99-100% chance 
> of success, and unlikely() is implemented as a 0-1% chance of success.
>
> As such, for our purposes, likely() and unlikely() should only be used 
> when a situation is [likely | unlikely] across all runtime 
> configurations.  So if you mark a branch unlikely() when it is hit often 
> by 1% of your users, that is an incorrect usage.
>
> The effects are probably most dramatic on older CPUs.  Repeatedly 
> hitting an unlikely() can cause a pipeline stall on every single access. 
>  Branch delay slots are filled improperly, with obvious implications.
>
> But on modern hardware, I would /guess/ that the effect of repeatedly 
> hitting an unlikely() would be mitigated by smarter branch prediction.
>
> We really need a GCC expert to answer this question in any more detail.

Aside from using branch constructs or hints that help the predictor
guess the right way... I think gcc will move unlikely paths right past
the end of the "likely" fastpath, so it can increase code size and be
somewhat suboptimal in terms of icache usage.

I don't know particularly why it would hurt a lot more when it goes
wrong than it helps when it goes right, though.

Also, I don't think I agree that it should be used where it is correct
for all users. We make rt_task unlikely in the scheduler, and I measured
that a very long time ago was IIRC good for nearly 5% pipe based context
switching peformance. Systems running a lot of rt tasks aren't going to
like it, but bugger them :)

--
SUSE Labs, Novell Inc.
-
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/