> By the way, my dht over replicate setup is indeed still showing a > performance boost with io-threads on the client, not just the server. > > Both cp -a /usr (<6 minutes instead of >8) and dd (sometimes the same, > sometimes >50MBps instead of <40MBps) tests come out significantly faster. > > It was stated that io-threads should no longer matter on the client-side in > 2.0, but that appears to be incorrect... The only reason I can imagine how iothreads could have helped, is by letting metadata ops overtake bulky IO ops at the transport. It should not have affected dd throughput in any way. Do you have flush-behind turned on in your write-behind? That might cause a few cases where loading io-threads gives an opportunity for further metadata calls to overtake the background write calls. Avati
