Need to set optimized parameters for installation on SSD disks

[Timur Nurlygayanov]

When we use SSD disks for installation of OpenStack/Ceph we need to use more optimized parameters to get normal performance.

1. Need to set disks scheduler to 'deadline' for SSD disks:
echo "deadline" >> /sys/block/<disk>/queue/scheduler

2. Mount it with the following options:
rw,noexec,inode64,logbsize=256k,delaylog,nodev,noatime,nodiratime,user_xattr,data=writeback

[Sergii Golovatiuk (sgolovatiuk)]

This bug is related to
https://bugs.launchpad.net/fuel/+bug/1318614

which causes high degradation because of 512b blocks instead of 4k. There should be a logic that will set the most optimal settings for different types of RAID or SSD disks.

[Dmitry Borodaenko (angdraug)]

Please use deadline scheduler instead of noop:
https://wiki.debian.org/SSDOptimization#Low-Latency_IO-Scheduler

[Timur Nurlygayanov (tnurlygayanov)]

Dmitry, thank you! Description was updated.

[Alexander Nevenchannyy (anevenchannyy)]

Since https://bugs.launchpad.net/fuel/+bug/1418126 we are now use deadline as default for all disks

[Mike Scherbakov (mihgen)]

Please consider if we can do required changes now since we've got IBP. It might be just simple logic as "if it's SSD, run these few lines / choose to run specific mount command". Please comment here.

[Aleksey Kasatkin (alekseyk-ru)]

It's feature-dependent and related BPs are moved to future release. So, it is moved to 8.0.

[Aleksey Kasatkin (alekseyk-ru)]

Seems to be partly solved by https://bugs.launchpad.net/fuel/+bug/1318614

[Alex Schultz (alex-schultz)]

MOS-Ceph please investigate what default configuration should be used.

[Alexey Shtokolov (ashtokolov)]

MOS-Ceph please give us an update about what default configuration should be used.

[Alexei Sheplyakov (asheplyakov)]

> 2. Mount it with the following options: rw,noexec,inode64,logbsize=256k,delaylog,nodev,noatime,nodiratime,user_xattr,data=writeback

The recommended filesystem for OSD storage is XFS. Unlike ext[34] it has a single journaling algorithm
(which is somewhat similar to data=writeback of ext[34]). Therefore there's no mount option(s) which
specify the journaling algorithm (which is what data=writeback|ordered|journal does for ext[34]]).

noexec,nodev have no effect on performance

delaylog is the XFS' default
inode64 is enabled on 64-bit systems by default
user_xattr is not a valid XFS mount option (extended attributes are always enabled)

[Piotr Skamruk (jell)]

AIR filesystem options was secondary in this issue. Main difference was done by IO scheduler for specific device.

If now XFS is default fs for OSD (as it should be, similar approach is usual suggestion for databases also) - second point is in fact out of date...