From: Christian Dietrich <stettberger@dokucode.de>
To: Hao Xu <haoxu@linux.alibaba.com>, Jens Axboe <axboe@kernel.dk>
Cc: io-uring@vger.kernel.org, Joseph Qi <joseph.qi@linux.alibaba.com>,
horst.schirmeier@tu-dresden.de,
"Franz-B. Tuneke" <franz-bernhard.tuneke@tu-dortmund.de>,
Hendrik Sieck <hendrik.sieck@tuhh.de>
Subject: Re: [POC RFC 0/3] support graph like dependent sqes
Date: Thu, 23 Dec 2021 11:06:17 +0100 [thread overview]
Message-ID: <s7by24bd49y.fsf@dokucode.de> (raw)
In-Reply-To: <20211214055734.61702-1-haoxu@linux.alibaba.com>
Hi everyone!
We experimented with the BPF patchset provided by Pavel a few months
ago. And I had the exact same question: How can we compare the benefits
and drawbacks of a more flexible io_uring implementation? In that
specific use case, I wanted to show that a flexible SQE-dependency
generation with BPF could outperform user-space SQE scheduling. From my
experience with BPF, I learned that it is quite hard to beat
io_uring+userspace, if there is enough parallelism in your IO jobs.
For this purpose, I've built a benchmark generator that is able to
produce random dependency graphs of various shapes (isolated nodes,
binary tree, parallel-dependency chains, random DAC) and different
scheduling backends (usual system-call backend, plain io_uring,
BPF-enhanced io_uring) and different workloads.
At this point, I didn't have the time to polish the generator and
publish it, but I put the current state into this git:
https://collaborating.tuhh.de/e-exk4/projects/syscall-graph-generator
After running:
./generate.sh
[sudo modprobe null_blk...]
./run.sh
./analyze.py
You get the following results (at least if you own my machine):
generator iouring syscall iouring_norm
graph action size
chain read 128 938.563366 2019.199010 46.48%
flat read 128 922.132673 2011.566337 45.84%
graph read 128 1129.017822 2021.905941 55.84%
rope read 128 2051.763366 2014.563366 101.85%
tree read 128 1049.427723 2015.254455 52.07%
For the userspace scheduler, I perform an offline analysis that finds
linear chains of operations that are not (anymore) dependent on other previous
unfinished results. These linear chains are then pushed into io_uring
with a SQE-link chain.
As I'm highly interested in this topic of pushing complex
IO-dependencies into the kernel space, I would be delighted to see how
your SQE-graph extension would compare against my rudimentary userspace
scheduler.
@Hao: Do you have a specific use case for your graph-like dependencies
in mind? If you need assistance with the generator, please feel
free to contact me.
chris
--
Prof. Dr.-Ing. Christian Dietrich
Operating System Group (E-EXK4)
Technische Universität Hamburg
Am Schwarzenberg-Campus 3 (E), 4.092
21073 Hamburg
eMail: christian.dietrich@tuhh.de
Tel: +49 40 42878 2188
WWW: https://osg.tuhh.de/
next prev parent reply other threads:[~2021-12-23 10:31 UTC|newest]
Thread overview: 15+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-12-14 5:57 [POC RFC 0/3] support graph like dependent sqes Hao Xu
2021-12-14 5:57 ` [PATCH 1/3] io_uring: add data structure for graph sqe feature Hao Xu
2021-12-14 5:57 ` [PATCH 2/3] io_uring: implement new sqe opcode to build graph like links Hao Xu
2021-12-14 5:57 ` [PATCH 3/3] io_uring: implement logic of IOSQE_GRAPH request Hao Xu
2021-12-14 15:21 ` [POC RFC 0/3] support graph like dependent sqes Pavel Begunkov
2021-12-14 16:53 ` Hao Xu
2021-12-14 18:16 ` Pavel Begunkov
2021-12-16 16:55 ` Hao Xu
2021-12-17 19:33 ` Pavel Begunkov
2021-12-18 6:57 ` Hao Xu
2021-12-21 16:19 ` Pavel Begunkov
2021-12-23 4:14 ` Hao Xu
2021-12-23 10:06 ` Christian Dietrich [this message]
2021-12-27 3:27 ` Hao Xu
2021-12-27 5:49 ` Christian Dietrich
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=s7by24bd49y.fsf@dokucode.de \
--to=stettberger@dokucode.de \
--cc=axboe@kernel.dk \
--cc=franz-bernhard.tuneke@tu-dortmund.de \
--cc=haoxu@linux.alibaba.com \
--cc=hendrik.sieck@tuhh.de \
--cc=horst.schirmeier@tu-dresden.de \
--cc=io-uring@vger.kernel.org \
--cc=joseph.qi@linux.alibaba.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).