At Tue, 05 Jun 2012 22:55:21 +0800, Liu Yuan wrote: > > On 06/05/2012 10:38 PM, MORITA Kazutaka wrote: > > > The below change to the qemu block driver would fix the problem. > > I think this kind of ordering should be done in the gateway node, > > though. > > > This will dramatically hurt the IO performance, at least for object > cache, I can see concurrent access to the same object, but NOT > overlapped at all. > > Most of disk driver default to 512K for maximum IO size, this means most > of chances we will get requests(split by OS driver and block layer) to > access the same 4M object. Concurrent accesses to the "same object" is not what I wanted to say. We can change the overlap check from "same object" to "same sector" easily. I'm just suggesting an approach to order requests in the client side. Thanks, Kazutaka |