[LU-13591] Enable lazytime mount for service nodes Created: 21/May/20 Updated: 24/May/20 Resolved: 22/May/20 |
|
| Status: | Closed |
| Project: | Lustre |
| Component/s: | None |
| Affects Version/s: | None |
| Fix Version/s: | None |
| Type: | Improvement | Priority: | Minor |
| Reporter: | Shaun Tancheff | Assignee: | Shaun Tancheff |
| Resolution: | Cannot Reproduce | Votes: | 0 |
| Labels: | None | ||
| Rank (Obsolete): | 9223372036854775807 |
| Description |
|
Using 'lazytime' when mounting MGS/MDS/OSS seems to offer performance advantages for 4.x and later kernels (where ldiskfs has support for lazytime). Enable this feature by default. |
| Comments |
| Comment by Andreas Dilger [ 22/May/20 ] |
|
Do you have any numbers to back this assertion? Since Lustre doesn't actually use/want the kernel updating the timestamps in the files, I don't think "lazytime" would make any difference there. It might make sense to add a "lazytime" mount option to the client, to minimize atime updates, if the clients don't already have "noatime" as a mount option. The MDS will already squash atime updates less than mdt.*.atime_diff seconds (default 60s) apart, but if the clients only send an atime update at most once a day that would be even better. |
| Comment by Shaun Tancheff [ 22/May/20 ] |
|
The early testing that showed some improvement did not show any advantage on subsequent runs. |
| Comment by Shaun Tancheff [ 22/May/20 ] |
|
Further benchmarking shows no clear advantage to lazytime on service nodes. |
| Comment by Andreas Dilger [ 24/May/20 ] |
|
I wouldn't be against implementing lazytime for the client, but that is up to you. |