On 05/23/2013 04:01 PM, Joseph Glanville wrote: > Yep, I have been reading that code. > > I think libaio would be more efficient when using faster backing > devices but it's probably not a problem due to how MD is structured. > If you have one vnode/md for each physical disk it should result in > good IO performance. > > I intend on testing sheepdog with some big 36 drive machines soon to > prove that theory though, try get feel for how well MD scales. Switch to aio isn't impossible and nonsense. Actually, I think aio model can work with current threading model similar to Linux's io scheduler that have one or several queues to run the requests and have io-threads to submit the io-requests to the queue(s). This might reduce the threads scheduling overhead and provide a more insight view of how requests can be executed more efficiently and provide the possibility to optimized the io scheduling on our own. Thanks, Yuan |