Re: [Lse-tech] Re: Patch and Performance of larger pipes

[Manfred S. <ma...@co...>]

Hubertus Franke wrote:
> 
> Manfred, here is some more performance measurement and comparision
> to round up some of the stuff we have been doing.
> It would be nice if you could make some comments. As usual raw numbers
> is not everything and there are other issues to consider.
> Comments ?
> 
I'm that much interested at SMP performance right now - most of the SMP
problems are bad interactions with the scheduler. The current scheduler
and the pipe implementation are "optimized" for each other, e.g.
wake_up_sync() is IIRC exclusively used by pipes.

> 
> UP
> --
>                                    %imp
> R.Size    W.Size    R.Compute W.Compute  2.4.9.pipe     Manfred
> ------    ------    -------------------     ----------   --------------
> 512        32k      500           0           0           1.04
> 512        32k        0         500       -0.91           7.72
> 
> 32k        512      500           0         1.09       -78.3
> 32k        512        0         500         0            0
>
I'll try to check that one ASAP. An uniprocessor performance degration
points to an implementation bug in my code.

--
	Manfred