[jira] Commented: (COUCHDB-754) Improve couch_file write performance

["Adam Kocoloski (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/COUCHDB-754?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12863593#action_12863593 ] 

Adam Kocoloski commented on COUCHDB-754:
----------------------------------------

DId a little more digging on the 8MB thing, found that it also occurs on 4K * 2048:

20> appending_microbench:test(4096, 2045).
94311
21> appending_microbench:test(4096, 2046).
279647

I'm guessing the reason the 2046th write takes ~180 ms is because the disk buffer filled up.  I checked the specs, the drive in my MacBook does have an 8MB buffer.  That's surprising to me, though -- I would have expected the data to stay cached by the kernel for much longer.  I never explicitly sync the file.

> Improve couch_file write performance
> ------------------------------------
>
>                 Key: COUCHDB-754
>                 URL: https://issues.apache.org/jira/browse/COUCHDB-754
>             Project: CouchDB
>          Issue Type: Improvement
>         Environment: some code might be platform-specific
>            Reporter: Adam Kocoloski
>             Fix For: 1.1
>
>         Attachments: cheaper-appending-v2.patch, cheaper-appending.patch
>
>
> I've got a number of possible enhancements to couch_file floating around in my head, wanted to write them down.
> * Use fdatasync instead of fsync.  Filipe posted a patch to the OTP file driver [1] that adds a new file:datasync/1 function.  I suspect that we won't see much of a performance gain from this switch because we append to the file and thus need to update the file metedata anyway.  On the other hand, I'm fairly certain fdatasync is always safe for our needs, so if it is ever more efficient we should use it.  Obviously, we'll need to fall back to file:sync/1 on platforms where the datasync function is not available.
> * Use file:pwrite/2 to batch together multiple outstanding write requests.  This is essentially Paul's zip_server [2].  In order to take full advantage of it we need to patch couch_btree to update nodes in parallel.  Currently there should only be 1 outstanding write request in a couch_file at a time, so it wouldn't help at all.
> * Open the file in append mode and stop seeking to eof in user space.  We never modify files (aside from truncating, which is rare enough to be handled separately), so perhaps it would help with performance if we let the kernel deal with the seek.  We'd still need a way to get the file size for the make_blocks function.  I'm wondering if file:read_file_info(Fd) is more efficient than file:position(Fd, eof) for this purpose.
> A caveat - I'm not sure if append-only files are compatible with the previous enhancement.  There is no file:write/2, and I have no idea how file:pwrite behaves on a file which is opened append-only.  Is the Pos ignored, or is it an error?  Will have to test.
> * Use O_DSYNC instead of fsync/fdatasync.  This one is inspired by antirez' recent blog post [3] and some historical discussions on pgsql-performance.  Basically, it seems that opening a file with O_DSYNC (or O_SYNC on Linux, which is currently the same thing) and doing all synchronous writes is reasonably fast.  Antirez' tests showed 250 µs delays for (tiny) synchronous writes, compared to 40 ms delays for fsync and fdatasync on his ext4 system.
> At the very least, this looks to be a compelling choice for file access when the server is running with delayed_commits = true.  We'd need to patch the OTP file driver again, and also investigate the cross-platform support.  In particular, I don't think it works on NFS.
> [1]: http://github.com/fdmanana/otp/tree/fdatasync
> [2]: http://github.com/davisp/zip_server
> [3]: http://antirez.com/post/fsync-different-thread-useless.html

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