openstack
/
charm-specs
Code Issues Proposed changes

Fixup formatting in ceph-autotune

This commit is contained in:
James Page 2016-06-07 18:09:07 +01:00
parent 3bde6a7a68
commit 113be9b536
1 changed files with 38 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

54
specs/newton/approved/ceph-autotune.rst
View File

@ -21,17 +21,27 @@ Ceph isn't as fast as it could be out of the box.
Problem Description
===================
Currently we don’t do any tuning of Ceph after installing it. There’s many little things that have to be tweaked to get ceph to be performant. There are many mailing list messages of users becoming frustrated with the slow performance of Ceph. The goal of this epic is to capture the low hanging fruit so everyone can have a better experience with Ceph.
Currently we don’t do any tuning of Ceph after installing it. There’s
many little things that have to be tweaked to get ceph to be performant.
There are many mailing list messages of users becoming frustrated with
the slow performance of Ceph. The goal of this epic is to capture the
low hanging fruit so everyone can have a better experience with Ceph.
Proposed Change
===============
I'm proposing several areas where easy improvements can be made. There's several areas that
can be improved such as HDD read ahead and max hardware sector settings. We can also
tune the sysctl network settings on 10 and 40Gb networks. Finally I've noticed that
IRQ alignment on multisocket motherboard isn't ideal. We can use numactl to pin the
IRQs for all the components in the IO path from network to osd to disk. Of all the changes
probably the IRQ pinning will be the most difficult.
I'm proposing several areas where easy improvements can be made.
There's several areas that can be improved such as HDD read ahead and
max hardware sector settings. We can also tune the sysctl network
settings on 10 and 40Gb networks.
Finally I've noticed that IRQ alignment on multisocket motherboard
isn't ideal. We can use numactl to pin the IRQs for all the components
in the IO path from network to osd to disk.
Of all the changes probably the IRQ pinning will be the most difficult.
Alternatives
------------
@ -51,23 +61,35 @@ Primary assignee:
Gerrit Topic
------------
Use Gerrit topic "performance-optimizations" for all patches related
to this spec.
.. code-block:: bash
git-review -t performance-optimizations
Work Items
----------
1. HDD Read ahead settings
* Often times the read ahead settings for hdd’s are too small for storage servers.
2. 10/40Gb sysctl settings
* 10Gb and 40Gb networks need special sysctl settings to improve their performance.
3. HDD max hardware sectors
* Making the max hardware sectors as large as possible allows bigger sequential writes and therefore better performance.
4. IRQ alignment for multisocket motherboards
* It has been shown on multi socket motherboards that Linux does a bad job of lining up IRQ’s and results in lots of swapping of cache memory between CPU’s. This significantly degrades Ceph’s performance.
1. HDD Read ahead settings: Often times the read ahead settings for
hdd’s are too small for storage servers.
2. 10/40Gb sysctl settings: 10Gb and 40Gb networks need special sysctl
settings to improve their performance.
3. HDD max hardware sectors: Making the max hardware sectors as large
as possible allows bigger sequential writes and therefore better
performance.
4. IRQ alignment for multisocket motherboards: It has been shown on
multi socket motherboards that Linux does a bad job of lining up
IRQ’s and results in lots of swapping of cache memory between CPU’s.
This significantly degrades Ceph’s performance.
Repositories
------------
https://github.com/openstack/charm-ceph-osd/
No new git repositories required for this work.
Documentation
-------------