On 02/20/2013 12:55 PM, Gregory Farnum wrote:
On Wed, Feb 20, 2013 at 10:36 AM, Mark Nelson <mark.nelson@xxxxxxxxxxx> wrote:
Hi Jonathan,
On 02/20/2013 12:28 PM, Jonathan Rudenberg wrote:
I'm currently planning a CEPH deployment, and we're looking at 36x4TB
drives per node. It seems like the recommended setup is an OSD per drive, is
this accurate? What is the recommended ratio of drives/OSDs per CPU core?
Would 12 cores be enough (3:1 ratio)?
Typically 1 drive per OSD is the way to go, but once you get up into the 36+
drives per node range there start becoming trade-offs (especially with
things like memory usage during recovery, etc). You may need to do some
testing to make sure that you don't end up hitting swap.
I've got a SC847a chassis we are using for testing at Inktank with 36 bays.
I'm using dual E5-2630ls and that seems to be working pretty well, but I
wouldn't go any slower than those chips. E5-2630s or 2640s might be a bit
better, but so far it looks like ivy bridge is fast enough that you can
fudge a bit on our "1ghz of CPU per OSD" guideline and get a pair of the
cheaper 6-core chips.
That 1GHz per daemon recommendation is based on recovery performance;
in general usage it'll often be much lower. I don't think you've done
much with recovery yet, so don't count on that ratio working out once
you do so!
-Greg
True enough. Having said that we've had those same recommendations for
a while now and per-clock performance has increased quite a bit since
the core2 and even Nehalem era. I'd be much more worried about memory
during recovery than CPU assuming that the CPU is specced relatively
close to our guidelines.
Mark
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
