osd disk latency at pve screen

[RobFantini]

Hello

at pve > osd I've a question on the latency columns.

We have all ssd osd. however some of the drives may not be ideal for ceph.


When a drive shows much greater latency then others - I assume they should be replaced.

Is there another test to verify that?

In mean time I'll start replacing the worst ones.

 

[Ashley]

Do you mean you have a mix of models across the OSD's?

 

[udo]

Hello

at pve > osd I've a question on the latency columns.

We have all ssd osd. however some of the drives may not be ideal for ceph.


When a drive shows much greater latency then others - I assume they should be replaced.

Is there another test to verify that?

In mean time I'll start replacing the worst ones.

Hi Rob,
the latency ist this, what "ceph osd perf" produce.

Higher values are also possible during deep-scrub... perhaps you should look, where the higher values came from before you switched the OSDs.

Udo

 

[RobFantini]

Do you mean you have a mix of models across the OSD's?

yes 3 diff models.

 

[RobFantini]

Hi Rob,
the latency ist this, what "ceph osd perf" produce.

Higher values are also possible during deep-scrub... perhaps you should look, where the higher values came from before you switched the OSDs.

Udo

thank you, I'll check that. however these disks - I've been recording the latency values - some generally always have higher latency then others.

 

[Ashley]

thank you, I'll check that. however these disks - I've been recording the latency values - some generally always have higher latency then others.

This can also be if you have particular hot data that is held on OSD's with these SSD's being the primary, however this will be hard to check.

However you can use some cli commands to dig deeper into the more busy disks to see what is making them busy, but if most of your busy OSD's point to a particular model of disk I would say that's your answer!

 

[RobFantini]

So would you say that it is best to use the same model on a ceph cluster?

 

[Ashley]

So would you say that it is best to use the same model on the ceph cluster?

Yes, as data is spread across multiple OSD's if each OSD has very different performance specifications it will cause some noticeable performance impact, aswel as causing the slow OSD to work harder, and the faster OSD not to be used to their full potential.

 

[RobFantini]

Thank you for the advise.