From: "Jeff V. Merkey" <jmerkey@vger.timpanogas.org>
To: linux-kernel@vger.kernel.org
Subject: queue_nr_requests needs to be selective
Date: Fri, 1 Mar 2002 13:22:54 -0700 [thread overview]
Message-ID: <20020301132254.A11528@vger.timpanogas.org> (raw)
Linus/Alan/Linux,
Performance numbers can be increased dramatically (> 300 MB/S)
by increasing queue_nr_requests in ll_rw_blk.c on large RAID
controllers that are hosting a lot of drives.
We have noticed that with the defaults of 64/128 for queue_nr_requests,
2.4.18 performance hits a wall at 230 MB/S system wide. We have gotten
this number to approach 300 MB/S across 4 adapters (for short periods
during cache flushing)
by putting 2 cards on the 33 Mhz bus and 2 cards on the 66 Mhx bus
in a Serverworks configuration running Dolphin's SCI adapters as
the interconnect to connect the adapters to an SCI clustering
fabric.
With queue_nr_requests set to 64/128, we have noticed via profile=2
that a lot of time is spent needlessly going to sleep in
__get_request_wait() with calling threads sleeping way more than is
needed. Since SCI does direct memory DMA,s we are at present pushing
238 MB/S from an SCI adapter into local cache and the cache is bursting
writes to the 3Ware adapters at very high data rates > 300 MB/S, but
only after I have hard coded 1024 into queue_nr_requests for the
disk queues in 2.4.18. We are at 34% utilization on a single
processor (no SMP) moving data at these rates with SCI.
What is really needed here is to allow queue_nr_requests to be
configurable on a per adapter/device basis for these high end
raid cards like 3Ware since in a RAID 0 configuration, 8 drives
are in essence a terabyte (1.3 terrabytes in our configuration)
and each adapter is showing up as a 1.3 TB device. 64/128
requests are simply not enough to get the full spectrum of
performance atainable with these cards.
These number needs to be increased. I cannot get any 2.4.18
system to get above 230 MB/S with 3ware RAID adapters with
the numbers set to 64/128 on these big devices. An array
of 8 x 160BG disks eats up these request queue free lists,
then the feeding threads spend most of their time asleep.
Respectfully submitted,
Jeff Merkey
next reply other threads:[~2002-03-01 20:09 UTC|newest]
Thread overview: 22+ messages / expand[flat|nested] mbox.gz Atom feed top
2002-03-01 20:22 Jeff V. Merkey [this message]
2002-03-01 20:43 ` Andrew Morton
2002-03-01 21:03 ` Alan Cox
2002-03-01 23:22 ` Jeff V. Merkey
2002-03-01 23:20 ` Jeff V. Merkey
2002-03-01 23:23 ` Andrew Morton
2002-03-02 0:27 ` Jeff V. Merkey
2002-03-02 0:49 ` Andrew Morton
2002-03-02 2:16 ` Jeff V. Merkey
2002-03-02 3:50 ` Andrew Morton
2002-03-02 4:34 ` Jeff V. Merkey
2002-03-02 7:33 ` Jeff V. Merkey
2002-03-02 9:10 ` Jens Axboe
2002-03-02 9:22 ` Andrew Morton
2002-03-04 9:09 ` Jens Axboe
2002-03-02 0:51 ` Mike Anderson
2002-03-02 4:39 ` Jeff V. Merkey
2002-03-02 5:59 ` Jeff V. Merkey
2002-03-02 6:01 ` Jeff V. Merkey
2002-03-02 6:16 ` Jeff V. Merkey
2002-03-04 7:16 ` Mike Anderson
2002-03-04 17:39 ` Jeff V. Merkey
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=20020301132254.A11528@vger.timpanogas.org \
--to=jmerkey@vger.timpanogas.org \
--cc=linux-kernel@vger.kernel.org \
/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
all inboxes | Powered by JetHome®