Le 06/12/10 18:58, Mark Riordan a écrit : > Interestingly, it didn't help to have any more than 2 read-aheads. > > Thanks for your help. > Mark,
My experience with the libssh user that requested that functionality (and sponsored for it) was that 10-20 readahead of 4kB requests was the optimal value. But it really depends on the latency of the link behind and of course, the performances of the disk behind. My feeling is that if some benchmark needs to send 1MB buffers to get decent performance on a LAN, something is wrong. Aris
