[performance] Mir spends 12%+ CPU time in get_hooked_symbol (libhybris-common) comparing strings

[Daniel van Vugt]

On the N4, Mir spends 12%+ CPU time in get_hooked_symbol (libhybris-common). Actually callgrind can't decide; it's somewhere between 12% and 46000% ;)

So it's big. Looking at the hybris code, there's a rather large bottleneck that's obvious: get_hooked_symbol does a linear search (many times) of a large list of strings using strcmp alone. That's why 22132 calls to get_hooked_symbol are yielding over 5 million calls to strcmp.

Upstream hybris appears to have been modified to use a cache/hashing instead of the ugly linear search we have.

[Daniel van Vugt (vanvugt)]

Hmm, consistent profile results seem to tell me that the offending code is possibly only called on startup. But if the startup code accounts for 12% of the CPU time of something I left rendering for several minutes, it's still worth fixing.

[Daniel van Vugt (vanvugt)]

Here's a patch where I've converted the linear search into a binary search. Performance is improved by a factor of approximately 10.

Even if this code is only touched on startup, it's definitely worth having because this represents most of the startup time of the Mir server.

[Ubuntu Foundations Team Bug Bot (crichton)]

The attachment "bsearch.patch" seems to be a patch. If it isn't, please remove the "patch" flag from the attachment, remove the "patch" tag, and if you are a member of the ~ubuntu-reviewers, unsubscribe the team.

[This is an automated message performed by a Launchpad user owned by ~brian-murray, for any issues please contact him.]

[Ricardo Salveti (rsalveti)]

Did you test the upstream implementation first to see if that is already enough for you? While I like the change you proposed, that would differ us from upstream, and the next rebase would mean dropping your change.

[Daniel van Vugt (vanvugt)]

I think hashing (upstream hybris) is a better idea and we should jump to that when possible. Although I just assumed we had no intention of rebasing given the significant diff. If we can then we should. I don't know what the performance diff to upstream is but upstream's hashing should possibly be even better than the bsearch.patch.

bsearch.patch is a short-term solution and long term we should rebase when possible.

[Launchpad Janitor (janitor)]

This bug was fixed in the package libhybris - 0.1.0+git20131207+e452e83-0ubuntu26

---------------
libhybris (0.1.0+git20131207+e452e83-0ubuntu26) utopic; urgency=medium

  [ Yu-Cheng Chou <email address hidden> ]
  * debian/control: fix typo in description (LP: #1259444)

  [ Daniel van Vugt ]
  * hooks: converting linear search into a binary search
    (LP: #1343198)
 -- Ricardo Salveti de Araujo <email address hidden> Thu, 14 Aug 2014 17:45:44 -0300