> Can you also measure the speed at which the client can write, when the > server is running each of the two kernel versions? If the client can > write twice as fast (for example) then it should be OK to use twice as > high a percentage of CPU time on the server, as the total CPU time > needed for a data transfer will be the same. >
I tested the following combinations with 2 GB writes: Server Client Debian testing kernel 2.6 Debian stable ~60 seconds Debian testing kernel 3.2 Debian stable ~60 seconds Debian testing kernel 2.6 Ubuntu 12.04.1 ~65 seconds Debian testing kernel 3.2 Ubuntu 12.04.1 ~67 seconds It seems the speed is comparable in all cases although the load differs. In general my load number looked a bit smaller this time, maybe because the host system was freshly booted. So I guess, I should collect some numbers on an actual machine tomorrow. I also did not have a Debian testing client yet. Do you know if the patch mentioned above also affect the client? --Karsten