shek@mullian.ee.mu.OZ.AU (Wilson SHEK) (11/13/89)
Recently, I conducted some round trip time measurements on UNIX DGRAM sockets in the INET domain and was puzzled by some of the results. The measurements were done between two SUN 3/50's (4 Mb main memory) under SunOS 4.0 over a 10Mbit ethernet. The setup of the test was as follows: one machine (at the sending end) sent an n-byte message using a DGRAM socket to the other machine (at the receiving end). As soon as the message arrived at the receiving end machine, it was sent back. The sending end machine would not send another message until the previous one was reflected back. This was repeated until 5000 round trips were completed. The total time required was recorded at the sending end machine using gettimeofday(). The following results were observed for n from 50 to 2000 with 50 byte increments. Message size Round trip time (total time/5000) (bytes) (ms) 50 5.37 . . . . RTT increases almost linearly . . 500 7.73 <- 550 6.99 <- an obvious drop at 512 octets 600 7.11 . . . . RTT increases almost linearly . . 1450 10.21 1500 13.08 <- 1550 11.97 <- another obvious drop at 1536 octets 1600 12.11 My questions are: 1. What caused the drops at 512 and 1536? 2. Why didn't the RTT drop significantly at 1024 bytes? (considering that the mbuf mechanism supporting the UNIX sockets was optimised for a page size of 1024 bytes) When the same test was done between two SUN3/60's, significant drop of RTT occurred at 1024,2048 (nothing special @ 512 and 1536). The tests were done when the network was virtually idle, therefore network traffic shouldn't have affected the results. Any help would be greatly appreciated. Thanks in advance. Wilson SHEK | ACSnet : shek@mullian.mu.oz Department of Elec Engineering | internet : shek@mullian.mu.oz.au University of Melbourne | uunet : uunet!munnari!mullian!shek