[Cluster-devel] [PATCH 2/2] gfs2: Extended attribute readahead optimization

[Andreas Gruenbacher]

On Thu, Nov 12, 2015 at 9:33 PM, Steven Whitehouse <swhiteho at redhat.com> wrote:
> That looks good to me, and I assume that it should be faster too?

I did some tests with a directory tree with 3439 directories and 51556
files in it.

In that tree, 47313 or 86% of the 54995 files and directories ended up
with readahead enabled, while 7682 didn't.

On a slow SATA magnetic disc, I compared the time required for running
"find", then "find | xargs stat", then "find | xargs getfattr". I
booted with selinux=0 for these measurements, so "find" and "find |
xargs stat" didn't use the xattrs.

The results are attached as a spreadsheet. They show that xattr
readahead can save quite some time (here, about -75%). The optimized
version of xattr readahead is about 3% faster than the simple version
in this test, so I'll leave it up to you whether that's worth it for
you.

It makes no difference whether the scheduler is "noop" or "deadline"
here; the "noop" scheduler still seems to merge adjacent I/O requests
(which makes sense).

Andreas
-------------- next part --------------
A non-text attachment was scrubbed...
Name: gfs2-readahead.ods
Type: application/vnd.oasis.opendocument.spreadsheet
Size: 11411 bytes
Desc: not available
URL: <http://listman.redhat.com/archives/cluster-devel/attachments/20151113/9323ca98/attachment.ods>