On Jun 4, 2007, at 8:17 PM, Simon Burge wrote:

> YAMAMOTO Takashi wrote:
>
>>> Anyway, I think this points to something funny going on with the  
>>> page
>>> loan code for pipes.  I don't know if this is exacerbated by running
>>> MP or what.  Obviously something is suboptimal with the default
>>> configuration choices here.
>>
>> IPIs involved in page flipping can make it slow on MP systems.
>> we should use some kind of cpu-local mapping.
>
> I just tried a UP NetBSD 3.1 system.  With kern.pipe.maxloankvasz at
> default (16MB) I get:
>
> -b 1k    74.04 MB/s
> -b 4k   112.38 MB/s
> -b 7k   166.79 MB/s
> -b 8k   166.87 MB/s
> -b 32k  230.22 MB/s
> -b 64k  226.93 MB/s
>
> and with kern.pipe.maxloankvasz set to 0 I get:
>
> -b 1k    78.22 MB/s
> -b 4k   205.99 MB/s
> -b 7k   239.62 MB/s
> -b 8k   351.07 MB/s
> -b 32k  630.09 MB/s
> -b 64k  790.75 MB/s
>
> So doesn't appear to be an IPI or general MP issue.

And what does PIPE_SOCKETPAIR (or whatever that kernel option is  
called) result in?

>
>
> Simon.

-- thorpej