[LU-11189] OSC flow control Created: 28/Jul/18 Updated: 26/Jan/23 |
|
| Status: | Open |
| Project: | Lustre |
| Component/s: | None |
| Affects Version/s: | None |
| Fix Version/s: | None |
| Type: | Improvement | Priority: | Minor |
| Reporter: | Jinshan Xiong | Assignee: | Zhenyu Xu |
| Resolution: | Unresolved | Votes: | 0 |
| Labels: | None | ||
| Issue Links: |
|
||||||||||||
| Rank (Obsolete): | 9223372036854775807 | ||||||||||||
| Description |
|
We're running into a problem that somehow OST devices are really slow to serve write requests. However, clients don't the situation so they would just keep sending requests, eventually this causes RPC timeout. This work is trying to relieve the problem by introducing a flow control on the OSC where it holds sending more requests as soon as it receives early reply. |
| Comments |
| Comment by Gerrit Updater [ 28/Jul/18 ] |
|
Jinshan Xiong (jinshan.xiong@gmail.com) uploaded a new patch: https://review.whamcloud.com/32895 |
| Comment by Andreas Dilger [ 28/Jul/18 ] |
|
On a related note, if there is heavy random read traffic on a large input file that is causing high IOPS and slowdown, you could consider to use "lfs ladvise willread /path/to/file" to prefetch the into RAM on the OST(s) it is located on. This requires that the input file is striped across enough OSTs to fit into RAM. |
| Comment by Andreas Dilger [ 23/Nov/20 ] |
|
Bobijam, could you please take a look at this patch and refresh it. I think we are still seeing similar problems under heavy load. |