Hi Sage and all, I am interested to do something in cache tiering. As we know, the two model caching and tiering are different ways to use SSD. In a tiering model, data is moved to the flash tier exclusively, meaning it never exists on both flash(cache tier) and hard disk(storage tier) drives. And in caching model, data is promoted to cache tier when read/write, and hot data exists in both cache tier and storage tier. Current Ceph cache tiering use cache model. If data is missed in cache tier, it needs to read from storage tier, and adding extra latency. The average latency is (latency of cahe tier)*percentage of cache hit + (latency of cache tier + lantency of storage tier)*percentage of cache miss. I am wondering how about we complete the feature with tiering model. The following is my raw thought: That is to say data in an image or object, hot data are put in cache tier, cold data are put in storage tier. Ceph tracks statistics of every object/chunks, and do promotion/demotion once every day or specified period. Meanwhile, combine librbd. First time client read data from an image, it can cache the real position(which pool data is) in client cache if not promotion/demotion ongoing. Later client can read data directly from cache tier or storage tier, no extra forward is needed. As a result, most of time we can get (latency of cache tier)*percentage of hot data + (latency of storage tier)*percentage of cold data. At the same time, I noticed that Sage talked something about rados tiering plan in following thread, could you please describe it more clearly, Sage? http://lists.ceph.com/pipermail/ceph-users-ceph.com/2017-February/016305.html -- Best wishes Lisa -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
