tree 154cfc0d8ff3b65f59b9052bcc41edaabf974063
parent 50069a5851323ba5def0e414a21e234345016870
author Doug Ledford <dledford@redhat.com> 1338506795 -0700
committer Linus Torvalds <torvalds@linux-foundation.org> 1338511771 -0700

ipc/mqueue: improve performance of send/recv

The existing implementation of the POSIX message queue send and recv
functions is, well, abysmal.  Even worse than abysmal.  I submitted a
patch to increase the maximum POSIX message queue limit to 65536 due to
customer needs, however, upon looking over the send/recv implementation, I
realized that my customer needs help with that too even if they don't know
it.  The basic problem is that, given the fairly typical use case scenario
for a large queue of queueing lots of messages all at the same priority (I
verified with my customer that this is indeed what their app does), the
msg_insert routine is basically a frikkin' bubble sort.  I mean, whoa,
that's *so* middle school.

OK, OK, to not slam the original author too much, I'm sure they didn't
envision a queue depth of 50,000+ messages.  No one would think that
moving elements in an array, one at a time, and dereferencing each pointer
in that array to check priority of the message being pointed too, again
one at a time, for 50,000+ times would be good.  So let's assume that, as
is typical, the users have found a way to break our code simply by using
it in a way we didn't envision.  Fair enough.

"So, just how broken is it?", you ask.  I wondered the same thing, so I
wrote an app to let me know.  It's my next patch.  It gave me some
interesting results.  Here's what it tested:

Interference with other apps - In continuous mode, the app just sits there
and hits a message queue forever, while you go do something productive on
another terminal using other CPUs.  You then measure how long it takes you
to do that something productive.  Then you restart the app in fake
continuous mode, and it sits in a tight loop on a CPU while you repeat
your tests.  The whole point of this is to keep one CPU tied up (so it
can't be used in your other work) but in one case tied up hitting the
mqueue code so we can see the effect of walking that 65,528 element array
one pointer at a time on the global CPU cache.  If it's bad, then it will
slow down your app on the other CPUs just by polluting cache mercilessly.
In the fake case, it will be in a tight loop, but not polluting cache.
Testing the mqueue subsystem directly - Here we just run a number of tests
to see how the mqueue subsystem performs under different conditions.  A
couple conditions are known to be worst case for the old system, and some
routines, so this tests all of them.

So, on to the results already:

Subsystem/Test                  Old                         New

Time to compile linux
kernel (make -j12 on a
6 core CPU)
  Running mqueue test     user 49m10.744s             user 45m26.294s
			   sys  5m51.924s              sys  4m59.894s
			 total 55m02.668s            total 50m26.188s

  Running fake test       user 45m32.686s             user 45m18.552s
                           sys  5m12.465s              sys  4m56.468s
                         total 50m45.151s            total 50m15.020s

  % slowdown from mqueue
    cache thrashing            ~8%                         ~.5%

Avg time to send/recv (in nanoseconds per message)
  when queue empty            305/288                    349/318
  when queue full (65528 messages)
    constant priority      526589/823                    362/314
    increasing priority    403105/916                    495/445
    decreasing priority     73420/594                    482/409
    random priority        280147/920                    546/436

Time to fill/drain queue (65528 messages, in seconds)
  constant priority         17.37/.12                    .13/.12
  increasing priority        4.14/.14                    .21/.18
  decreasing priority       12.93/.13                    .21/.18
  random priority            8.88/.16                    .22/.17

So, I think the results speak for themselves.  It's possible this
implementation could be improved by cacheing at least one priority level
in the node tree (that would bring the queue empty performance more in
line with the old implementation), but this works and is *so* much better
than what we had, especially for the common case of a single priority in
use, that further refinements can be in follow on patches.

[akpm@linux-foundation.org: fix typo in comment, remove stray semicolon]
[levinsasha928@gmail.com: use correct gfp flags in msg_insert]
Signed-off-by: Doug Ledford <dledford@redhat.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Manfred Spraul <manfred@colorfullife.com>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Sasha Levin <levinsasha928@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>