On Wed, 3 Dec 2014 15:44:31 -0500 Trond Myklebust <trond.myklebust@xxxxxxxxxxxxxxx> wrote: > On Wed, Dec 3, 2014 at 3:21 PM, Jeff Layton <jeff.layton@xxxxxxxxxxxxxxx> wrote: > > On Wed, 3 Dec 2014 14:59:43 -0500 > > Trond Myklebust <trond.myklebust@xxxxxxxxxxxxxxx> wrote: > > > >> On Wed, Dec 3, 2014 at 2:20 PM, Jeff Layton <jeff.layton@xxxxxxxxxxxxxxx> wrote: > >> > On Wed, 3 Dec 2014 14:08:01 -0500 > >> > Trond Myklebust <trond.myklebust@xxxxxxxxxxxxxxx> wrote: > >> >> Which workqueue are you using? Since the receive code is non-blocking, > >> >> I'd expect you might be able to use rpciod, for the initial socket > >> >> reads, but you wouldn't want to use that for the actual knfsd > >> >> processing. > >> >> > >> > > >> > I'm using the same (nfsd) workqueue for everything. The workqueue > >> > isn't really the bottleneck though, it's the work_struct. > >> > > >> > Basically, the problem is that the work_struct in the svc_xprt was > >> > remaining busy for far too long. So, even though the XPT_BUSY bit had > >> > cleared, the work wouldn't get picked up again until the previous > >> > workqueue job had returned. > >> > > >> > With the change I made today, I just added a new work_struct to > >> > svc_rqst and queue that to the same workqueue to do svc_process as soon > >> > as the receive is done. That means though that each RPC ends up waiting > >> > in the queue twice (once to do the receive and once to process the > >> > RPC), and I think that's probably the reason for the performance delta. > >> > >> Why would the queuing latency still be significant now? > >> > > > > That, I'm not clear on yet and that may not be why this is slower. But, > > I was seeing slightly faster performance with reads before I made > > today's changes. If changing how these jobs get queued doesn't help the > > performance, then I'll have to look elsewhere... > > Do you have a good method for measuring that latency? If the queuing > latency turns out to depend on the execution latency for each job, > then perhaps running the message receives on a separate low latency > queue could help (hence the suggestion to use rpciod). > I was using ftrace with the sunrpc:* and workqueue:* tracepoints, and had a simple perl script to postprocess the trace info to figure out average/min/max latency. I don't think the queueing latency is that significant per-se, but I think the best thing is to avoid making multiple trips through the workqueue per RPC if we can help it. I tested and pushed a newer patchset to my repo last night that does that (at least if there's already a svc_rqst available when the xprt needs servicing). It seems to be pretty close speed-wise to the thread-based code. The next step is to test this out on something larger-scale. I'm hoping to get access to just such a test rig soon. Once we have some results from that, I think I'll have a much better idea of how viable this approach is and where other potential bottlenecks might be. Thanks! -- Jeff Layton <jlayton@xxxxxxxxxxxxxxx> -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html