On 04/12/2012 11:30 PM, Christoph Hellwig wrote: > On Thu, Apr 12, 2012 at 10:53:22PM +0800, Liu Yuan wrote: >> From: Liu Yuan <tailai.ly at taobao.com> >> >> Most of the time one VM would issue multiple requests in one >> go, so it would be useful to let users to decide how many workers >> are useful if we have more than several VMs in single sheep node. >> >> In a large set nodes of cluster, every single node will get multiple >> concurrent recovery IO requests, so it would be useful to have it as >> configurable. >> >> default 4 workers for both and maximum 128. > > I've seen in my benchmarking that's you'd want at least 64 to even > saturate setup with single sata disks per daemon and a few dozens > daemons, which means the default probably should be in that range. > Hi Christoph, Thanks for your test numbers, I didn't expect this large threads to run not-so-fast disk busy. How many cpus do you have on your box? I think this number has a great dependency with the host available cpu number. I am not sure, I'd love to set 64 as default since our servers has a considerable cpu cores, but for the boxes with limited cpus (Say, only 2), is 64 threads okay for them too(Do not waste memory resource for unnecessary task_struct)? Thanks, Yuan |