On 10/6/21 09:23, Zakhar Kirpichenko wrote:
Hi,
Indeed, that's a lot of CPU and RAM, the idea was to put sufficient
resources in case we want to expand the nodes with more storage and do
EC. I guess having excessive resources shouldn't hurt performance? :-)
That was also my take. Untill an (Oracle) DBA explained to me that in
cases were you don't use / need the cache, it can hurt performance as it
still costs resources to manage the cache. When using linux (with buffer
cache enabled) RAM probably won't go to waste.
My modus operandi is "no better kill than overkill" ... but on the other
hand, you seem to want to optimize performance, not necessarily fill
both CPU sockets. For a Ceph storage node I would want to have more OSDs
instead of more memory / CPU (as long as there is plenty). Not sure if
you run into any NUMA issues here, but a single socket system with say
anywhere between 16-64 cores seem like a better fit. If you're going to
attach a JBOD and add way more disks, then sure, you might need all
these resources. But in that case I would still wonder if more nodes
with less resources would be a better fit. Just thinking out loud here,
although indirectly related to performance, not directly answering your
disappointing I/O performance thread.
Gr. Stefan
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
