From: "Martin J. Bligh" <mbligh@aracnet.com>
To: Ingo Molnar <mingo@elte.hu>, Erich Focht <efocht@ess.nec.de>
Cc: Michael Hohnbaum <hohnbaum@us.ibm.com>,
Matthew Dobson <colpatch@us.ibm.com>,
Christoph Hellwig <hch@infradead.org>,
Robert Love <rml@tech9.net>, Andrew Theurer <habanero@us.ibm.com>,
Linus Torvalds <torvalds@transmeta.com>,
linux-kernel <linux-kernel@vger.kernel.org>,
lse-tech <lse-tech@lists.sourceforge.net>
Subject: Re: [patch] sched-2.5.59-A2
Date: Mon, 20 Jan 2003 08:23:20 -0800 [thread overview]
Message-ID: <1372810000.1043079799@titus> (raw)
In-Reply-To: <Pine.LNX.4.44.0301201022540.2585-100000@localhost.localdomain>
> kernelbench is the kind of benchmark that is most sensitive to over-eager
> global balancing, and since the 2.5.59 ->nr_balanced logic produced the
> best results, it clearly shows it's not over-eager.
Careful ... what shows well on one machine may not on another - this depends
heavily on the NUMA ratio - for our machine the nr_balanced logic in 59 is
still over-agressive (20:1 NUMA ratio). For low-ratio machines it may work
fine. It actually works best for us when it's switched off altogether I
think (which is obviously not a good solution).
But there's more than one dimension to tune here - we can tune both the
frequency and the level of imbalance required. I had good results specifying
a minimum imbalance of > 4 between the current and busiest nodes before
balancing. Reason (2 nodes, 4 CPUs each): If I have 4 tasks on one node,
and 8 on another, that's still one or two per cpu in all cases whatever
I do (well, provided I'm not stupid enough to make anything idle). So
at that point, I just want lowest task thrash.
Moving tasks between nodes is really expensive, and shouldn't be done
lightly - the only thing the background busy rebalancer should be fixing
is significant long-term imbalances. It would be nice if we also chose
the task with the smallest RSS to migrate, I think that's a worthwhile
optimisation (we'll need to make sure we use realistic benchmarks with
a mix of different task sizes). Working out which ones have the smallest
"on-node RSS - off-node RSS" is another step after that ...
> hackbench is one that is quite sensitive to under-balancing.
> Ie. trying to maximize both will lead us to a good balance.
I'll try to do some hackbench runs on NUMA-Q as well.
Just to add something else to the mix, there's another important factor
as well as the NUMA ratio - the size of the interconnect cache vs the
size of the task migrated. The interconnect cache on the NUMA-Q is 32Mb,
our newer machine has a much lower NUMA ratio, but effectively a much
smaller cache as well. NUMA ratios are often expresssed in terms of
latency, but there's a bandwidth consideration too. Hyperthreading will
want something different again.
I think we definitely need to tune this on a per-arch basis. There's no
way that one-size-fits-all is going to fit a situation as complex as this
(though we can definitely learn from each other's analysis).
M.
next prev parent reply other threads:[~2003-01-20 16:14 UTC|newest]
Thread overview: 96+ messages / expand[flat|nested] mbox.gz Atom feed top
2003-01-09 23:54 Minature NUMA scheduler Martin J. Bligh
2003-01-10 5:36 ` [Lse-tech] " Michael Hohnbaum
2003-01-10 16:34 ` Erich Focht
2003-01-10 16:57 ` Martin J. Bligh
2003-01-12 23:35 ` Erich Focht
2003-01-12 23:55 ` NUMA scheduler 2nd approach Erich Focht
2003-01-13 8:02 ` Christoph Hellwig
2003-01-13 11:32 ` Erich Focht
2003-01-13 15:26 ` [Lse-tech] " Christoph Hellwig
2003-01-13 15:46 ` Erich Focht
2003-01-13 19:03 ` Michael Hohnbaum
2003-01-14 1:23 ` Michael Hohnbaum
2003-01-14 4:45 ` [Lse-tech] " Andrew Theurer
2003-01-14 4:56 ` Martin J. Bligh
2003-01-14 11:14 ` Erich Focht
2003-01-14 15:55 ` [PATCH 2.5.58] new NUMA scheduler Erich Focht
2003-01-14 16:07 ` [Lse-tech] " Christoph Hellwig
2003-01-14 16:23 ` [PATCH 2.5.58] new NUMA scheduler: fix Erich Focht
2003-01-14 16:43 ` Erich Focht
2003-01-14 19:02 ` Michael Hohnbaum
2003-01-14 21:56 ` [Lse-tech] " Michael Hohnbaum
2003-01-15 15:10 ` Erich Focht
2003-01-16 0:14 ` Michael Hohnbaum
2003-01-16 6:05 ` Martin J. Bligh
2003-01-16 16:47 ` Erich Focht
2003-01-16 18:07 ` Robert Love
2003-01-16 18:48 ` Martin J. Bligh
2003-01-16 19:07 ` Ingo Molnar
2003-01-16 18:59 ` Martin J. Bligh
2003-01-16 19:10 ` Christoph Hellwig
2003-01-16 19:44 ` Ingo Molnar
2003-01-16 19:43 ` Martin J. Bligh
2003-01-16 20:19 ` Ingo Molnar
2003-01-16 20:29 ` [Lse-tech] " Rick Lindsley
2003-01-16 23:31 ` Martin J. Bligh
2003-01-17 7:23 ` Ingo Molnar
2003-01-17 8:47 ` [patch] sched-2.5.59-A2 Ingo Molnar
2003-01-17 14:35 ` Erich Focht
2003-01-17 15:11 ` Ingo Molnar
2003-01-17 15:30 ` Erich Focht
2003-01-17 16:58 ` Martin J. Bligh
2003-01-18 20:54 ` NUMA sched -> pooling scheduler (inc HT) Martin J. Bligh
2003-01-18 21:34 ` [Lse-tech] " Martin J. Bligh
2003-01-19 0:13 ` Andrew Theurer
2003-01-17 18:19 ` [patch] sched-2.5.59-A2 Michael Hohnbaum
2003-01-18 7:08 ` William Lee Irwin III
2003-01-18 8:12 ` Martin J. Bligh
2003-01-18 8:16 ` William Lee Irwin III
2003-01-19 4:22 ` William Lee Irwin III
2003-01-17 17:21 ` Martin J. Bligh
2003-01-17 17:23 ` Martin J. Bligh
2003-01-17 18:11 ` Erich Focht
2003-01-17 19:04 ` Martin J. Bligh
2003-01-17 19:26 ` [Lse-tech] " Martin J. Bligh
2003-01-18 0:13 ` Michael Hohnbaum
2003-01-18 13:31 ` [patch] tunable rebalance rates for sched-2.5.59-B0 Erich Focht
2003-01-18 23:09 ` [patch] sched-2.5.59-A2 Erich Focht
2003-01-20 9:28 ` Ingo Molnar
2003-01-20 12:07 ` Erich Focht
2003-01-20 16:56 ` Ingo Molnar
2003-01-20 17:04 ` Ingo Molnar
2003-01-20 17:10 ` Martin J. Bligh
2003-01-20 17:24 ` Ingo Molnar
2003-01-20 19:13 ` Andrew Theurer
2003-01-20 19:33 ` Martin J. Bligh
2003-01-20 19:52 ` Andrew Theurer
2003-01-20 19:52 ` Martin J. Bligh
2003-01-20 21:18 ` [patch] HT scheduler, sched-2.5.59-D7 Ingo Molnar
2003-01-20 22:28 ` Andrew Morton
2003-01-21 1:11 ` Michael Hohnbaum
2003-01-22 3:15 ` Michael Hohnbaum
2003-01-22 16:41 ` Andrew Theurer
2003-01-22 16:17 ` Martin J. Bligh
2003-01-22 16:20 ` Andrew Theurer
2003-01-22 16:35 ` Michael Hohnbaum
2003-02-03 18:23 ` [patch] HT scheduler, sched-2.5.59-E2 Ingo Molnar
2003-02-03 20:47 ` Robert Love
2003-02-04 9:31 ` Erich Focht
2003-01-20 17:04 ` [patch] sched-2.5.59-A2 Martin J. Bligh
2003-01-21 17:44 ` Erich Focht
2003-01-20 16:23 ` Martin J. Bligh [this message]
2003-01-20 16:59 ` Ingo Molnar
2003-01-17 23:09 ` Matthew Dobson
2003-01-16 23:45 ` [PATCH 2.5.58] new NUMA scheduler: fix Michael Hohnbaum
2003-01-17 11:10 ` Erich Focht
2003-01-17 14:07 ` Ingo Molnar
2003-01-16 19:44 ` John Bradford
2003-01-14 16:51 ` Christoph Hellwig
2003-01-15 0:05 ` Michael Hohnbaum
2003-01-15 7:47 ` Martin J. Bligh
2003-01-14 5:50 ` [Lse-tech] Re: NUMA scheduler 2nd approach Michael Hohnbaum
2003-01-14 16:52 ` Andrew Theurer
2003-01-14 15:13 ` Erich Focht
2003-01-14 10:56 ` Erich Focht
2003-01-11 14:43 ` [Lse-tech] Minature NUMA scheduler Bill Davidsen
2003-01-12 23:24 ` Erich Focht
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=1372810000.1043079799@titus \
--to=mbligh@aracnet.com \
--cc=colpatch@us.ibm.com \
--cc=efocht@ess.nec.de \
--cc=habanero@us.ibm.com \
--cc=hch@infradead.org \
--cc=hohnbaum@us.ibm.com \
--cc=linux-kernel@vger.kernel.org \
--cc=lse-tech@lists.sourceforge.net \
--cc=mingo@elte.hu \
--cc=rml@tech9.net \
--cc=torvalds@transmeta.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).