https://bugzilla.mcs.anl.gov/swift/show_bug.cgi?id=690 --- Comment #29 from David Kelly <[email protected]> 2012-02-21 15:13:19 --- I have finished re-running the socket tests while recording all packets with tcpdump/wireshark. I wrote a perl script to extract the window sizes of each packet to see what is actually happening for each test. There are a few things that jump out while looking at the data. When no socket options are specified, it looks exactly like what we thought - as the file transfer progresses, the window size increases. It seems plausible that this could be causing some issues. http://skynetdevelopment.com/network_tests/auto.png When I set SO_SNDBUF and SO_RCVBUF, the kernel seems to be honoring that request. Window sizes are no larger than double the value I request. There is no tcp window scaling happening here at all, but the transfer still freezes. There also seems to be a correlation between the window size and the total number of packets successfully sent. When SNDBUF and RCVBUF are 1024, the least number of packets transfer successfully before freezing. As I increase those values, the number of packets successfully transferred increases too. This pattern continues until a SNDBUF and RCVBUF reach 64K. http://skynetdevelopment.com/network_tests/under64k.png When SNDBUF and RCVBUF values are between 64K and 4096K, the window size stays at 64K. It seems to have no effect within this range. http://skynetdevelopment.com/network_tests/over64k.png I am passing along this information now. -- Configure bugmail: https://bugzilla.mcs.anl.gov/swift/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are watching all bug changes.