On Mon, Sep 19, 2016 at 2:20 PM, Jeff Darcy <jdarcy@xxxxxxxxxx> wrote: > FWIW, I did some further experiments. Disabling mem-pool entirely (in favor of plain malloc/free) brought run time down to 3:35, vs. 2:57 for the exact same thing without multiplexing. Somehow we're still not managing contention very well at this kind of thread count, but the clues and opportunities are becoming less obvious. I wonder if we are spending more time in io-threads. Does setting idle-time in io-threads to 1 help with anything? It might be useful to add instrumentation subsequently to dump statistics (number of fops serviced, time spent in servicing) per thread. Having the same visibility for our epoll threads would also be useful. I think the answer is no - but does profiling yield anything useful at this point? -Vijay _______________________________________________ Gluster-devel mailing list Gluster-devel@xxxxxxxxxxx http://www.gluster.org/mailman/listinfo/gluster-devel
