Merge lp://staging/~xfactor973/charms/trusty/ceph-osd/sysctl-perf into lp://staging/~openstack-charmers-archive/charms/trusty/ceph-osd/trunk
Status: | Superseded |
---|---|
Proposed branch: | lp://staging/~xfactor973/charms/trusty/ceph-osd/sysctl-perf |
Merge into: | lp://staging/~openstack-charmers-archive/charms/trusty/ceph-osd/trunk |
Diff against target: |
550 lines (+307/-19) 10 files modified
.bzrignore (+3/-0) .testr.conf (+8/-0) config.yaml (+10/-1) hooks/ceph.py (+203/-7) hooks/ceph_hooks.py (+28/-10) hooks/install (+1/-1) requirements.txt (+11/-0) templates/ceph.conf (+5/-0) test-requirements.txt (+9/-0) tox.ini (+29/-0) |
To merge this branch: | bzr merge lp://staging/~xfactor973/charms/trusty/ceph-osd/sysctl-perf |
Related bugs: |
Reviewer | Review Type | Date Requested | Status |
---|---|---|---|
James Page | Needs Fixing | ||
Review via email:
|
This proposal has been superseded by a proposal from 2016-02-18.
Description of the change
This branch adds some HDD and network adapter sysctl tuning. Lots of potential performance has been left on the table with our out of the box ceph setup. This aims to recapture some of that.
For the HDD portion: Linux has set very conservative settings when it comes to HDDs. While this is great for desktop responsiveness it isn't so kind to storage servers. I have exposed a configuration option if an administrator knows they have a RAID card with a lot of cache onboard. In a future patch set I will add some SSD tuning to take advantage of its unique characteristics.
Testing this has proven challenging. While it doesn't break anything on AWS that doesn't really prove anything either. If anyone has access to physical hardware to test this I would really appreciate it.
Unmerged revisions
- 56. By Chris Holcombe
-
Removed enum dependency
- 55. By Chris Holcombe
-
Pull in the tox files from next
- 54. By Chris Holcombe
-
Persist the changes for reboots. Also catch IOError in all the posix operations
- 53. By Chris Holcombe
-
Address the lint issues and try to solve the python-enum34 package missing with the unit tests
- 52. By Chris Holcombe
-
Change over to using network link speed as the key to sysctl tuning instead of the driver name. This is a more generic approach because most 10Gb links need roughly the same set of sysctls
- 51. By Chris Holcombe
-
Lint spacing
- 50. By Chris Holcombe
-
Add a status notification to the network adapter tuning section
- 49. By Chris Holcombe
-
Added possible network adapter sysctl tuning
- 48. By Chris Holcombe
-
Add a config flag and a function to tune the read and write sectors for each osd block device
- 47. By Chris Holcombe
-
Sysctl and performance tweaks for ceph
See comment on network interfaces. I'd also like to see some validation on varied hardware (SSD / HDD) that this doesn't break things either way.