On Mon 28-03-11 10:44:45, Wu Fengguang wrote:

Hi Jan,

On Sat, Mar 26, 2011 at 07:05:44AM +0800, Jan Kara wrote:

quoted

  Hello Fengguang,

On Fri 25-03-11 21:44:11, Wu Fengguang wrote:

quoted

On Wed, Mar 23, 2011 at 05:43:14AM +0800, Jan Kara wrote:

quoted

  Hello Fengguang,

On Fri 18-03-11 22:30:01, Wu Fengguang wrote:

quoted

On Wed, Mar 09, 2011 at 06:31:10AM +0800, Jan Kara wrote:

quoted

  Hello,

  I'm posting second version of my IO-less balance_dirty_pages() patches. This
is alternative approach to Fengguang's patches - much simpler I believe (only
300 lines added) - but obviously I does not provide so sophisticated control.

Well, it may be too early to claim "simplicity" as an advantage, until
you achieve the following performance/feature comparability (most of
them are not optional ones). AFAICS this work is kind of heavy lifting
that will consume a lot of time and attention. You'd better find some
more fundamental needs before go on the reworking.

(1)  latency
(2)  fairness
(3)  smoothness
(4)  scalability
(5)  per-task IO controller
(6)  per-cgroup IO controller (TBD)
(7)  free combinations of per-task/per-cgroup and bandwidth/priority controllers
(8)  think time compensation
(9)  backed by both theory and tests
(10) adapt pause time up on 100+ dirtiers
(11) adapt pause time down on low dirty pages 
(12) adapt to new dirty threshold/goal
(13) safeguard against dirty exceeding
(14) safeguard against device queue underflow

  I think this is a misunderstanding of my goals ;). My main goal is to
explore, how far we can get with a relatively simple approach to IO-less
balance_dirty_pages(). I guess what I have is better than the current
balance_dirty_pages() but it sure does not even try to provide all the
features you try to provide.

OK.

quoted

I'm thinking about tweaking ratelimiting logic to reduce latencies in some
tests, possibly add compensation when we waited for too long in
balance_dirty_pages() (e.g. because of bumpy IO completion) but that's
about it...

Basically I do this so that we can compare and decide whether what my
simple approach offers is OK or whether we want some more complex solution
like your patches...

Yeah, now both results are on the website. Let's see whether they are
acceptable for others.

  Yes. BTW, I think we'll discuss this at LSF so it would be beneficial if
we both prepared a fairly short explanation of our algorithm and some
summary of the measured results. I think it would be good to keep each of
us below 5 minutes so that we don't bore the audience - people will ask for
details where they are interested... What do you think?

That looks good, however I'm not able to attend LSF this year, would
you help show my slides?

  Ah, that's a pity :(. If you send me a few slides I can show them, that's
no problem. I'll also try to understand your patches in enough detail so
that I can answer possible questinons but author is always the best to
present his work :).
 

quoted

I'll try to run also your patches on my setup to see how they work :) V6
from your website is the latest version, isn't it?

Thank you. You can run 
http://git.kernel.org/?p=linux/kernel/git/wfg/writeback.git;a=shortlog;h=refs/heads/dirty-throttling-v6
or 
http://www.kernel.org/pub/linux/kernel/people/wfg/writeback/dirty-throttling-v6/dirty-throttling-v6-2.6.38-rc6.patch
whatever convenient for you.

If you are ready with v3, I can also help test it out and do some
comparison on the results.

  I have done a couple of smaller fixes but I don't expect them to affect
performance in the loads we use. But I'll send you the patches when I
implement some significant change (but for that I need to reproduce the
latencies you sometimes see first...).

									Honza
-- 
Jan Kara [off-list ref]
SUSE Labs, CR

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>