Re: RFC on io-stalls patch

[Jens Axboe]

On Sat, Jul 12 2003, Chris Mason wrote:
> On Sat, 2003-07-12 at 03:37, Jens Axboe wrote:
> 
> > > I believe the new way provides better overall read performance in the
> > > presence of lots of writes.
> > 
> > I fail to see the logic in that. Reads are now treated fairly wrt
> > writes, but it would be really easy to let writes consume the entire
> > capacity of the queue (be it all the requests, or just going oversized).
> > 
> > I think the oversized logic is flawed right now, and should only apply
> > to writes. Always let reads get through. And don't let writes consume
> > the last 1/8th of the requests, or something like that at least. I'll
> > try and do a patch for pre4.
> 
> If we don't apply oversized checks to reads, what keeps a big streaming
> reader from starving out all the writes?

It's just so much easier to full the queue with writes than with reads.

> The current patch provides a relatively fixed amount of work to get a
> request, and I don't think we should allow that to be bypassed.  We
> might want to add a special case for synchronous reads (like bread), via
> a b_state bit that tells the block layer an immediate unplug is coming
> soon.  That way the block layer can ignore the oversized checks, grant a
> request and unplug right away, hopefully lowering the total number of
> unplugs the synchronous reader has to wait through.
> 
> Anyway, if you've got doubts about the current patch, I'd be happy to
> run a specific benchmark you think will show the bad read
> characteristics.

No I don't have anything specific, it just seems like a bad heuristic to
get rid of. I can try and do some testing tomorrow. I do feel strongly
that we should at least make sure to reserve a few requests for reads
exclusively, even if you don't agree with the oversized check. Anything
else really contradicts all the io testing we have done the past years
that shows how important it is to get a read in ASAP. And doing that in
the middle of 2.4.22-pre is a mistake imo, if you don't have numbers to
show that it doesn't matter for the quick service of reads.

-- 
Jens Axboe