FWIW, I'd be for trying to take periodic samples of actual IO happening
on the drive during operation. You can get a much better idea of
latency and throughput characteristics across different IO sizes over
time (though you will need to account for varying levels of concurrency
at the device level). The goal would be to start building a model of
drive throughput/latency under different scenarios. It would be a lot
of work, but could be useful for a wide variety of things.
Mark
On 9/8/22 08:09, Sridhar Seshasayee wrote:
Frank, Sven,
I totally agree with your thoughts on performing a proper benchmark study.
The
osd bench results we obtained when compared with Fio results didn't vary
wildly
and therefore the osd bench was adopted considering the minimal impact to
the
osd startup time. But this apparently is not always the case.
In case the osd bench throws up unrealistic numbers, the recommendation
would be to run a proper benchmark using reliable tools so that allocations
for
different types of operations are accurate and scheduling operations
function
correctly.
In the immediate term we will improve the mclock documentation around
this so that users are aware what to do in case they encounter such a
situation.
We will also explore options to determine IOPS capacity more reliably.
Thank you for your feedback!
-Sridhar
_______________________________________________
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
