On Wed, Feb 27, 2019 at 12:06:32PM +0100, Solene Rapenne wrote: > Copy a file over NFS (3 times) > > Before patch: 20.51 / 19.74 / 19.74 seconds > After patch: 14.48 / 15.03 / 14.49 seconds > > It gets 25% faster!
Nice :) > Ping -c100 -i0.2 on local network > > Before patch: 2.398/3.869/10.290/1.347 ms > After patch: 2.470/3.245/5.608/0.492 ms Ping times can vary widely and also depend on concurrent transmissions made by other devices. But the trend looks good, and it is expected that we're now causing less latency on the channel ourselves. > 3 differents file get from NFS (once because after it in cache, not > sure if test is relevant or reliable): > > Before patch: file1 11.39 / file2 12.21 / file3 6.47 seconds > After patch: file1 12.44 / file2 12.50 / file3 7.26 seconds > > This test show a small slowdown while other tests show a big > improvement. It should be faster in general, but we'll still flip RTS on if the MiRA heuristic believes it makes sense and that can lead to periods of time during which performance drops again. We might still need to tweak this heuristic in the future. There are a lot of details the paper does not explain. We'll have to figure out what works best in practice. Thanks for testing! The diff has been committed earlier today.