On 07/02/2013 10:29 PM, Sage Weil wrote:
Hi Wido!
On Tue, 2 Jul 2013, Wido den Hollander wrote:
Something in the back of my mind keeps saying that there were plans to
implement read caching in librbd, but I haven't been able to find any
reference about that.
In the tracker however I wasn't able to find anything, so it has to be me.
Any plans or ideas on this? A simple idea was that librbd could use a mmap
where it caches RADOS objects and purges them on a write.
This cache could then live on an SSD on a hypervisor where it could speed up
read operations for Virtual Machines in OpenStack or CloudStack envs.
Something like FS-Cache for librbd.
This has come up in a few different conversations over the last few weeks.
This would be a good thing to discuss in a blueprint for emperor. In my
mind there are two different use case classes that we want to cover:
Ah, great! We blueprint in the Wiki would be the best start I guess with
a issue in the tracker.
- persistent caching (e.g., a local SSD)
Something like that, doesn't have to be persistent imho, but it would be
a bonus.
- shared per-host cache (e.g., many qemu instances with a shared cache of
the clone base images)
Ack, that would be even more awesome!
Wido
My hope is that we can build a simple caching interface that plugs into
ObjectCacher (the per-process in-memory cache) with multiple
implementations (shared memory, files on disk) to kill both birds with one
stone. The trick will be around cache invalidation. Even if we
somewhat punt on that, though, a simple approach would be usable for the
read-only gold image cache (say, in memory) case.
sage
--
Wido den Hollander
42on B.V.
Phone: +31 (0)20 700 9902
Skype: contact42on
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
