[LU-9252] osd: Cannot really disable cache on OSS Created: 24/Mar/17 Updated: 05/Aug/20 Resolved: 31/Jan/19 |
|
| Status: | Resolved |
| Project: | Lustre |
| Component/s: | None |
| Affects Version/s: | None |
| Fix Version/s: | None |
| Type: | Improvement | Priority: | Minor |
| Reporter: | CEA | Assignee: | Peter Jones |
| Resolution: | Duplicate | Votes: | 0 |
| Labels: | performance | ||
| Issue Links: |
|
||||||||
| Severity: | 3 | ||||||||
| Rank (Obsolete): | 9223372036854775807 | ||||||||
| Description |
|
When deactivating caching on an OSS through the osd proc files: read_cache_enable and writethrough_cache_enable, it only simulates Lustre's behaviour without cache. It would be interesting though, for Lustre's OSSes to be able to allocate their memory outside of the pagecache for it seems to be a performance issues when using SSD disks: it creates a CPU overload and 65% of CPU ressources are spent to allocate/free memory from the pagecache. |
| Comments |
| Comment by Andreas Dilger [ 24/Mar/17 ] |
|
To clarify your concern, the issue is that even with read/write cache disabled the OSS is still adding the pages to the mapping for the inode and then dropping it after the read/write is complete. It should be possible to change this caching behaviour for osd-ldiskfs to preallocate IO buffers per thread, in a manner similar to what Lustre 1.8 did, but I don't think it would be possible to do this for osd-zfs. |
| Comment by Andreas Dilger [ 31/Jan/19 ] |
|
This was fixed via |