On Tue, Nov 15, 2011 at 07:26:55PM +0800, Liu Yuan wrote: > On 11/15/2011 06:43 PM, Christoph Hellwig wrote: > > > It's not going to help with read performance indeed. In my benchmarks > > so read performance wasn't a major issue, mostly because I always had > > a copy of all objects stored on the machine where qemu runs. > > > So this is not true of most regular cases. Current consistent-hashing > will sprinkle objects all over the nodes of the cluster and the nodes > location are subjective to changing because of data rebalance at > membership change. No local copy kept. Yes - I just meant to say that so far I didn't trigger read performance issue yet - they onbviously exist if data is far away. > > If you care about read performance having the objects locally is what > > you need - adding a config tweak that makes sure to keep a local copy > > of objects read at least once might be a good idea. > > > > > What do you mean exactly by 'config tweak'? In a setup where all nodes are close to each other and the qemu instance migrating data to be local on a read isn't going to help much, so behaviour to do this should not be unconditionally. > We might use readahead not only for performance, but also it improves > the availability on node basis. For e.g, if we read-ahead the whole > image locally, this node can work in a controlled state that it doesn't > care about other nodes at all, whether them be died or live in some sense. Yes. I'm very interested in a model like that. > > The proposal linked above probaby isn't too benefical for write > > performance, given that you only start pushing things to the network > > once the flush routine is called, and thus use a lot of bandwith in > > the latency critical flush roundtrip. Sending unstable write requests > > to all nodes ASAP, and only doing the final sync on flush will get > > much better performance. > > > Yes, that will cause latency issue. But we are concerned about reducing > network traffic too, how is your plan going to reduce the network traffic? Maybe I haven't understood the linked model yet, but I don't see how it reduces network traffic either. |