Mailing-List: contact dev-help@couchdb.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@couchdb.apache.org
Message-ID: <30552339.29271272944335459.JavaMail.jira@thor>
Date: Mon, 3 May 2010 23:38:55 -0400 (EDT)
From: "Adam Kocoloski (JIRA)" <jira@apache.org>
To: dev@couchdb.apache.org
Subject: [jira] Commented: (COUCHDB-738) more efficient DB compaction (fewer
 seeks)
In-Reply-To: <18200813.141191271341491073.JavaMail.jira@thor>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


    [ https://issues.apache.org/jira/browse/COUCHDB-738?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12863646#action_12863646 ] 

Adam Kocoloski commented on COUCHDB-738:
----------------------------------------

Ok, here's a first pass.  I ran the usual perf.py script which loads 1000 'simple' and 1000 'unsimple' docs into a DB and computes the MegaView and SimpleView.  I modified the script to update each document 100 times instead of only once.  Here are the numbers for trunk:

DB size: 1.1G pre-compaction, 12MB post

MegaView
Elapsed time: 7.20 (s)
Elapsed time: 5.65 (s)
Elapsed time: 5.13 (s)

SimpleView
Elapsed time: 0.75 (s)
Elapsed time: 1.04 (s)
Elapsed time: 0.74 (s)

Indexing times after compaction
Elapsed time: 4.91 (s)
Elapsed time: 0.74 (s)

Note that only the first measurement included a full load_docs step.  For the latter measurements I just deleted and rebuilt the index files (updating the docs 100 times took several minutes, and I didn't feel like waiting).  Here are the equivalent numbers with the patch:

DB size: 1.5G pre-compaction, 17MB post

MegaView
Elapsed time: 9.03 (s)
Elapsed time: 8.22 (s)
Elapsed time: 5.92 (s)

SimpleView
Elapsed time: 0.99 (s)
Elapsed time: 0.94 (s)
Elapsed time: 0.91 (s)

After compaction
Elapsed time: 5.81 (s)
Elapsed time: 0.90 (s)

There seems to be quite a bit of variance in these numbers.  I usually take the minimum when benchmarking on my laptop, although due to my skipping the load_docs step the minimum is probably a little too good (the seq_tree is fully cached).  Anyway, if I stick with that methodology I'd say that for DBs with 100 edits/document we see

MegaView: 15-18% slower with patch 
SimpleView: 22-23% slower with patch

The database is 35-40% larger, too.

I might try loading a 500 edits / document DB tonight.  I may also look at eprof to see if there are simple optimizations for this use case.  Is the number of edit branches significant?  So far I haven't been introducting any conflicts in the test.

> more efficient DB compaction (fewer seeks)
> ------------------------------------------
>
>                 Key: COUCHDB-738
>                 URL: https://issues.apache.org/jira/browse/COUCHDB-738
>             Project: CouchDB
>          Issue Type: Improvement
>          Components: Database Core
>    Affects Versions: 0.9.2, 0.10.1, 0.11
>            Reporter: Adam Kocoloski
>            Assignee: Adam Kocoloski
>             Fix For: 1.1
>
>         Attachments: 738-efficient-compaction-v1.patch
>
>
> CouchDB's database compaction algorithm walks the by_seq btree, then does a lookup in the by_id btree for every document in the database.  It does this because the #full_doc_info{} record with the full revision tree is only stored in the by_id tree.  I'm proposing instead to store duplicate copies of #full_doc_info{} in both trees, and to have the compactor use the by_seq tree exclusively.  The net effect is significantly fewer calls to pread(), and an compaction IO pattern where reads tend to be clustered close to each other in the file.
> If the by_id tree is fully cached, or if the id tree nodes are located near the seq tree nodes, the performance improvement is small but noticeable (~10% in some simple tests).  On the other hand, in the worst-case scenario of randomly-generated docids and a database much larger than main memory the improvement is huge.  Joe Williams did some simple benchmarks with a 50k document, 600 MB database on a 256MB VPS.  The compaction time for that DB dropped from 15m to 2m20s, so more than 6x faster.
> Storing the #full_doc_info{} in the seq tree also allows for some similar optimizations in the replicator.
> This patch might have downsides when documents have a large number of edits.  These include an increase in the size of the database and slower view indexing.  I expect both to be small effects.
> The patch can be applied directly to trunk@934272.  Existing DBs are still readable, new updates will be written in the new format, and databases can be fully upgraded by compacting.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.