Was comparing xfstests genetic/070 with and without the compounding patches in for-next. Between my VMs I saw a reduction from 85-90s to 45-49s reduction in wall time for execution repeated over several tries. A little less that the previous microbenchmarks but still showing very promising numbers. Nearly cutting the running time for generic/070 in half. Two things here help performance. We first eliminate the inter-command gaps, i.e. the delays between the client receives one reply and until it issues the next command. This is the cost of bouncing in and out of the event loop. Secondly we also get the reduction in number of roundtrips. For very low networks I suspect that the former may be a significant factor for the improvement. For remote servers I expect the latter to be totally dominating for where the performance improvements come from. Would be really cool to see some numbers when accessing a server tens of ms away. Any help testing, finding and fixing bugs welcome. This has been touching a lot of very hairy code so there are bound to be issues that we need to find/fix. regards ronnie sahlberg
