https://bz.apache.org/bugzilla/show_bug.cgi?id=63916
--- Comment #18 from Mark Thomas <ma...@apache.org> --- Tomcat is calling write until 0 bytes are written before adding to the poller so that isn't what is going on here. (In reply to Remy Maucherat from comment #12) > By default on my Linux the socket send buffer size is 2304. For NIO to have > the same "performance" and behavior running the test [as seen with the debug > logs] as NIO2, the send buffer size needs to be exactly >= to 4737. On my Linux box the default send buffer size is 1313280. With the stand-alone test, I see slowness if the buffer size < 4537 and no slowness when it is >=4537. It does seem that performance is related to how many times we have to call select(). With a larger buffer size, the network write seems to be more efficient so there are far fewer calls to select(). I'd expect the stand-alone test to be writing faster than Tomcat and therefore need a larger buffer to avoid the slowness. I want to add use of a executor for the writes to see if that gives more realistic results. -- You are receiving this mail because: You are the assignee for the bug. --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@tomcat.apache.org For additional commands, e-mail: dev-h...@tomcat.apache.org