On Monday, May 4, 2015, Christian Balzer <ch...@gol.com> wrote:

> On Mon, 13 Apr 2015 10:39:57 +0530 Sanjoy Dasgupta wrote:
>
> > Hi!
> >
> > This is an often discussed and clarified topic, but Reason why I am
> > asking is because
> >
> > If We use a RAID controller with Lot of Cache (FBWC) and Configure each
> > Drive as Single Drive RAID0, then  Write to disks will benefit by using
> > FBWC and accelerate I/O performance. Is this correct assumption ?
> >
> In the case of Ceph, the journal writes (assuming journals on the HDDs,
> not separate SSDs) will benefit from this indeed.


Each of my OSD nodes has 7 2TB disks on one RAID card with 1GB of FBWC with
BBU.

I get pretty good performance. With 6 of the above osd nodes and no
replication, using rados bench, I can get around 7500 4k write iops and it
can peak to over 10k iops.

The write throughput at 4MB is around 1000 MB/s, It will peak to around
1309 MB/s.




> > Also, it indeed helps, what are the downside of using RAID Controller
> > with Cache in above manner (Apart from Cost of RAID controller) ?
> >
> If you have too much money or existing HW, knock yourself out. ^.^
>
> Aside from the cost, the cache should be battery backed, additional costs
> and maintenance issues.
> Setting up RAID0 drives can be painful (megacli must die), and adds
> another step to OSD replacement/deployment.
> The resulting drive may or may not have all SMART features exposed and
> (not applicable in your case) won't support TRIM.
>
> Lastly the cache tends to be small when shared with many HDDs and there is
> also competition over it by reads and writes, but that's something to keep
> in mind, not a disadvantage per se.


Yes, all of the above.

I think it is worth the money.

I thought I wouldn't need SSD journals, and the raid cards were cheaper
than a few good SSDs.

Due to my odd workload, I end up with all of my writes being below 256k in
size.  I have to use iscsi proxy nodes to connect to my VMware hosts. This
limits my VM's throughput to around 200MB/s which isn't fast enough for my
application.  I'm planning on building new osd nodes with the same raid
cards and SSD journals to help further coalesce the small writes.



> Christian
> --
> Christian Balzer        Network/Systems Engineer
> ch...@gol.com <javascript:;>        Global OnLine Japan/Fusion
> Communications
> http://www.gol.com/
> _______________________________________________
> ceph-users mailing list
> ceph-users@lists.ceph.com <javascript:;>
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>
_______________________________________________
ceph-users mailing list
ceph-users@lists.ceph.com
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com

Reply via email to