Result for a single ls on a directory with 10k directories inside (16*3
volume)
*
*
Configuration
Plain volume
Parallel-readdir
Proposed Solution
Single Dir ls (Seconds)
-----
135
32.744
It is showing 321% improvements.
Regards
Rafi KC
On 27/05/20 11:26 am, RAFI KC wrote:
Hi All,
I have been working on POC to improve readdirp performance
improvement. At the end of the experiment, The results are showing
promising result in performance, overall there is a 104% improvement
for full filesystem crawl compared to the existing solution. Here is
the short test numbers. The tests were carried out in 16*3 setup with
1.5 Million dentries (Both files and dir). The system also contains
some empty directories. *In the result the proposed solution is 287%
faster than the plane volume and 104% faster than the parallel-readdir
based solution.*
*
*
Configuration
Plain volume
Parallel-readdir
Proposed Solution
FS Crawl Time in Seconds
16497.523
8717.872
4261.401
In short, the basic idea behind the proposal is the efficient managing
of readdir buffer in gluster along with prefetching the dentries for
intelligent switch-over to the next buffer. The detailed problem
description, deign description and results are available in the
doc.https://docs.google.com/document/d/10z4T5Sd_-wCFrmDrzyQtlWOGLang1_g17wO8VUxSiJ8/edit
If anybody can help with the testing on a different kind of workloads,
I would be very happy to assist. If wanted to test the patch and run a
performance test on your setup, I could help with back-porting the
patch to the version of your choice.
https://review.gluster.org/24469
https://review.gluster.org/24470
Regards
Rafi KC
________
Community Meeting Calendar:
Schedule -
Every 2nd and 4th Tuesday at 14:30 IST / 09:00 UTC
Bridge: https://bluejeans.com/441850968
Gluster-users mailing list
[email protected]
https://lists.gluster.org/mailman/listinfo/gluster-users
________
Community Meeting Calendar:
Schedule -
Every 2nd and 4th Tuesday at 14:30 IST / 09:00 UTC
Bridge: https://bluejeans.com/441850968
Gluster-users mailing list
[email protected]
https://lists.gluster.org/mailman/listinfo/gluster-users