Bonjour, During benchmarking[1], I noticed that intensive RBD client I/O can hurt performances and latency. That is, if clients read and write as much as they can, the total throughput may be negatively impacted and the latency of individual reads and writes will increase. If the clients throttle their I/O, it improves both the throughput and the latency.
Ideally the RBD QoS settings[2] could be tuned so that clients do not have to worry about throttling at all. They would read and write as fast as they can and Ceph would make sure it does not create a situation that would significantly degrade performances. There are about two dozen parameters and setting up experiments to figure out how they interact with each other is non trivial. I would be interested in any real world tutorials / recommendation / feedback / blog posts on the matter. Cheers P.S. disclaimer: I did not research and an answer such as "Read The Fine Blog/Manual" with a URL would be most welcome :-) [1] https://wiki.softwareheritage.org/wiki/Preliminary_Object_Storage_Benchmarks [2] https://docs.ceph.com/en/pacific/rbd/rbd-config-ref/#qos-settings -- Loïc Dachary, Artisan Logiciel Libre
_______________________________________________ ceph-users mailing list -- [email protected] To unsubscribe send an email to [email protected]
