From: David Smith <dsmith@redhat.com>
To: "Frank Ch. Eigler" <fche@redhat.com>
Cc: systemtap@sources.redhat.com, utrace-devel@redhat.com
Subject: Re: syscall tracing overheads: utrace vs. kprobes
Date: Tue, 28 Apr 2009 18:19:00 -0000 [thread overview]
Message-ID: <49F748C3.10307@redhat.com> (raw)
In-Reply-To: <20090428170117.GA30001@redhat.com>
Frank Ch. Eigler wrote:
> Hi -
>
> In a few contexts, it comes up as to whether it is faster to probe
> process syscalls with kprobes or with something higher level such as
> utrace. (There are other hypothetical options too (per-syscall
> tracepoints) that could be measured this way in the future.)
These scenarios are a bit wrong:
> Now we compare these scenarios:
>
> # stap -e 'probe never {}' -t --vp 00001 -c a.out
>
> Here, no actual probing occurs so we get a measurement of the plain
> uninstrumented run time of ten million close(2)s.
The above one is fine.
> # stap -e 'probe process.syscall {}' -t --vp 00001 -c a.out
>
> Here, we intercept sys_close with a kprobe. If the system is not too
> busy, we should pick up only the close(2)s coming from a.out, though a
> few close(2)'s executed by other processes may show up.
>
> # stap -e 'probe syscall.close {}' -t --vp 00001 -c a.out
>
> Here, we intercept all a.out's syscalls with utrace. Other processes
> are not affected at all, but other syscalls by a.out would be --
> though in our test, there are hardly any of those.
These 2 are swapped: the 'process.syscall' probe is a utrace-based
probe and the 'syscall.close' probe is a kprobe-based probe.
Note that in the results, the description and probe types matched correctly.
> Some typical results on my 2.66GHz 2*Xeon5150 machine runnin Fedora 9 -
> 2.6.27.12:
>
> never:
> Pass 5: run completed in 740usr/3310sys/4155real ms.
>
> kprobe:
> probe syscall.close (<input>:1:1), hits: 10000028, cycles: 176min/202avg/3632max
> Pass 5: run completed in 750usr/9320sys/10193real ms.
>
> utrace:
> probe process.syscall (<input>:1:1), hits: 10000025, cycles: 176min/209avg/184392max
> Pass 5: run completed in 1670usr/6860sys/8645real ms.
>
> So utrace added 4.5 seconds, and kprobes added 6.0 seconds to the
> uninstrumented 4.1 second run time. But wait: we should subtract the
> time taken by the probe handler itself: 200ish cycles at 2.66 GHz,
> which is about 0.75 seconds. So the overheads are approximately:
>
> never: n/a
> kprobe: 5.2 seconds => 0.52 us per hit
> utrace: 3.6 seconds => 0.36 us per hit
>
>
> Note that these are microbenchmarks that represent an ideal case
> compared to a larger run, since they probably fit comfily inside
> caches. They probably also undercount the probe handler's run time.
--
David Smith
dsmith@redhat.com
Red Hat
http://www.redhat.com
256.217.0141 (direct)
256.837.0057 (fax)
prev parent reply other threads:[~2009-04-28 18:19 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2009-04-28 17:01 Frank Ch. Eigler
2009-04-28 18:10 ` Roland McGrath
2009-04-28 18:16 ` Frank Ch. Eigler
2009-04-28 18:45 ` Roland McGrath
2009-04-28 18:19 ` David Smith [this message]
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=49F748C3.10307@redhat.com \
--to=dsmith@redhat.com \
--cc=fche@redhat.com \
--cc=systemtap@sources.redhat.com \
--cc=utrace-devel@redhat.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 read-only IMAP folder(s) and NNTP newsgroup(s).