Hi,
to have a fair test you need to replicate the power loss scenarios ceph
does cover and you are currently not:
No memory caches in the os or an the disk are allowed to be used, ceph
has to ensure that an object written is actually written, even if a node
of your cluster explodes right at that moment and you get a
simultaneously black out no data is allowed to be lost.
I have seen multiple GB/s ssds (without hardware power loss) degrade to
single digit MB/s due to this.
Ceph greatly scales out with a lot of disks and nodes, but in smaller
systems you will definitely notice the lower per disk performance, also
latency can be an issue, eg when doing blockwise single threaded writes
over a blocking file api.
Greetings,
Kai
On 1/16/22 14:31, Behzad Khoshbakhti wrote:
Hi Marc,
Thanks for your prompt response.
We have test the direct random write for the disk (without Ceph) and it is
200 MB/s. Wonder why we got 80MB/s from Ceph.
Your help is much appreciated.
Regards,
Behzad
On Sun, Jan 16, 2022 at 11:56 AM Marc <Marc@xxxxxxxxxxxxxxxxx> wrote:
Detailed (somehow) problem description:
Disk size: 1.2 TB
Ceph version: Pacific
Block size: 4 MB
Operation: Sequential write
Replication factor: 1
Direct disk performance: 245 MB/s
Ceph controlled disk performance: 80 MB/s
you are comparing sequential io against random. You should consider that
ceph writes to the drive in a 'random manner'
Sorry for asking this dummy question as I know there are numerous
parameters affecting the performance.
Yes this is written about everywhere on this list.
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
