Mailing-List: contact dev-help@couchdb.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@couchdb.apache.org
Date: Tue, 19 Apr 2011 22:10:05 +0000 (UTC)
From: "Randall Leeds (JIRA)" <jira@apache.org>
To: dev@couchdb.apache.org
Message-ID: 
 <1385091262.68299.1303251005773.JavaMail.tomcat@hel.zones.apache.org>
In-Reply-To: 
 <1096411398.65711.1303163885757.JavaMail.tomcat@hel.zones.apache.org>
Subject: [jira] [Updated] (COUCHDB-1130) binary optimization in couch_file
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


     [ https://issues.apache.org/jira/browse/COUCHDB-1130?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Randall Leeds updated COUCHDB-1130:
-----------------------------------

    Attachment: prof_bin_opt

Here's a runnable escript that exercises the original version, the patched version I uploaded before, and a third version that tries to avoid the match context stuff I talked about.

Since the docs I linked are super confusing I'm not going to try to dig too far into explanations by my findings are:

The original version is slowest.
The first patch is faster.
The third version (in the escript) is even faster.

We're talking 1us or 2us per call.

If there's a difference in memory consumption I can't tell from any of the output or any of the erlang:system_info or erlang:memory functions. But then, I didn't try to mess with the tracers or profiles to see what memory consumption is like mid-execution, so maybe all the resources used are already freed at the end of the test.

Anyway. Do with this what you will. It's a small perf gain but if it's reproducible it's a pretty critical code path to optimize so I say go for it?

I also tried flipping the order of the tests around just in case that had any affect, but the results are consistent.

> binary optimization in couch_file
> ---------------------------------
>
>                 Key: COUCHDB-1130
>                 URL: https://issues.apache.org/jira/browse/COUCHDB-1130
>             Project: CouchDB
>          Issue Type: Improvement
>          Components: Database Core
>    Affects Versions: 1.0.2
>            Reporter: Randall Leeds
>            Priority: Minor
>         Attachments: 0001-refactor-remove_block_prefixes-2-for-optimization.patch, prof_bin_opt
>
>
> I've had this patch sitting around since January and kept forgetting to file the ticket. Hurray spring cleaning.
> Just for fun I ran erlc with +bin_opt_info, which gives information about how the Erlang VM can optimize creating binary objects.
> What follows is the commit message from my patch.
> Even if I'm wrong about the last point, it can't hurt.
> What think you all?
> -------------------
> Running erlc with +bin_opt_info gives an INFO message stating that "matching
> anything else but a plain variable to the left of a binary pattern will prevent
> delayed sub binary optimization; SUGGEST changing argument order"
> I guess matching 0 is triggering this. If I understand correctly, this change
> will allow the compiler to skip creating a sub-binary that starts at the block
> boundary in the third clause and delay creation until we strip the leading byte
> in the 0 clause. This means one less 1-byte binary every time we read across a
> block boundary.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira