Re: [PATCH] slabify iocontext + request_queue

Re: [PATCH] slabify iocontext + request_queue

[Christoph Hellwig]

this just went in:

>  From: Jens Axboe <axboe@suse.de>
> 
> Move both request_queue and io_context allocation to a slab cache.
> 
> This is mainly a space-saving exercise.  Some setups have a lot of disks
> and the kmalloc rounding-up can consume significant amounts of memory.

While I agree on the io_context part, slabifying request_queue is a space
waste on most machines out there.  The averange desktop has less than a
handfull of these, and even for smaller servers it doesn't exactly look
like a gain.

Re: [PATCH] slabify iocontext + request_queue

[Jens Axboe]

On Sat, May 15 2004, Christoph Hellwig wrote:
> this just went in:
> 
> >  From: Jens Axboe <axboe@suse.de>
> > 
> > Move both request_queue and io_context allocation to a slab cache.
> > 
> > This is mainly a space-saving exercise.  Some setups have a lot of disks
> > and the kmalloc rounding-up can consume significant amounts of memory.
> 
> While I agree on the io_context part, slabifying request_queue is a space
> waste on most machines out there.  The averange desktop has less than a
> handfull of these, and even for smaller servers it doesn't exactly look
> like a gain.

See the thread last week on queue congestion threshold calculations,
there were some numbers in there.

-- 
Jens Axboe

Re: [PATCH] slabify iocontext + request_queue

[Christoph Hellwig]

On Sat, May 15, 2004 at 07:30:04PM +0200, Jens Axboe wrote:
> > While I agree on the io_context part, slabifying request_queue is a space
> > waste on most machines out there.  The averange desktop has less than a
> > handfull of these, and even for smaller servers it doesn't exactly look
> > like a gain.
> 
> See the thread last week on queue congestion threshold calculations,
> there were some numbers in there.

Maybe my math is completely off, but with slab you'd need a page at least,
the kmem_cache_t and maybe a kmem_bufctl_t

So for the usual two or three queue desktops we went from 1024 or 1536
to 4096 + N.

Cutoff point is at aproximately 9 queues which I think most machines running
linux won't reach.

Anyway, not that this is really important, I think we just need to question
all this silent bloating a little..