Re: RFC on io-stalls patch

[Jens Axboe <axboe@suse.de>]

On Wed, Jul 16 2003, Andrea Arcangeli wrote:
> On Tue, Jul 15, 2003 at 01:27:37PM +0200, Jens Axboe wrote:
> > On Tue, Jul 15 2003, Alan Cox wrote:
> > > On Maw, 2003-07-15 at 06:26, Jens Axboe wrote:
> > > > Sorry, but I think that is nonsense. This is the way we have always
> > > > worked. You just have to maintain a decent queue length still (like we
> > > > always have in 2.4) and there are no problems.
> > > 
> > > The memory pinning problem is still real - and always has been. It shows up
> > > best not on IDE disks but large slow media like magneto opticals where you
> > > can queue lots of I/O but you get 500K/second
> > 
> > Not the same thing. On slow media, like dvd-ram, what causes the problem
> > is that you can dirty basically all of the RAM in the system. That has
> > nothing to do with memory pinned in the request queue.
> 
> you can trivially bound the amount of dirty memory to nearly 0% with the
> bdflush sysctl. And the overkill size of the queue until pre3 could be
> an huge VM overhead compared to the dirty memory on lowmem boxes,
> example a 32/64M machine. So I disagree it's only a mistake of write
> throttling that gives problems on slow media.
> 
> Infact I tend to think the biggest problem for slow media in 2.4 is the
> lack of per spindle pdflush.

Well it's a combined problem. Threshold too high on dirty memory,
someone doing a read well get stuck flushing out as well.

-- 
Jens Axboe