Bug #1122
closed
Added by Sage Weil almost 13 years ago.
Updated almost 13 years ago.
Description
Many people now have noticed that sequential read performance is slower than writes. Is this simply a matter of adjusting readahead defaults, or is something else going on?
- Translation missing: en.field_story_points set to 5
- Translation missing: en.field_position set to 685
- Translation missing: en.field_position deleted (
683)
- Translation missing: en.field_position set to 17
- Target version changed from v0.30 to v0.31
- Translation missing: en.field_story_points changed from 5 to 8
- Translation missing: en.field_position deleted (
27)
- Translation missing: en.field_position set to 27
- Translation missing: en.field_position deleted (
26)
- Translation missing: en.field_position set to 701
- Target version changed from v0.31 to v0.32
- Translation missing: en.field_position deleted (
704)
- Translation missing: en.field_position set to 11
- Target version changed from v0.32 to v0.33
- Translation missing: en.field_position deleted (
41)
- Translation missing: en.field_position set to 1
- Subject changed from slow read performance to kclient: async readahead
- Translation missing: en.field_position deleted (
1)
- Translation missing: en.field_position set to 1
- Assignee set to Sage Weil
- Status changed from New to Resolved
I'm seeing significantly better performance on sepia. E.g. for 8 osds/1 client, kclient reads saturate the interface (around 117 MB/sec) while cfuse reads get about 70 MB/sec.
Before the changes I see 91 MB/sec (with the same cluster/nodes).
dd if=/dev/zero of=foo bs=1M count=10000
for f in `seq 1 10` ; do dd if=foo of=/dev/null bs=1M ; done
Also available in: Atom
PDF