mills@HUEY.UDEL.EDU.UUCP (01/19/87)
Folks, I boogied with a fuzzball far away (Stuttgart, W. Germany) from here today and collected the data for another scatter diagram. The results, using ICMP Echo messages uniformly distributed in length from 40 to 576 octets, proved rather gruesome. The minimum roundtrip delay was a few hundred milliseconds, and the maximum almost ten seconds, while the regression line had a y intercept of 1845, 576-octet intercept of 3927 and slope of 2212 bps (all in milliseconds). There was a surprisingly large cliff between the singel-packet and multi-packet regimes at about 127 octets of over one second, reminiscent of the "old" ARPANET protocols studied in RFC-889 and thought to be reduced in recent years. In fact, I conclude based on these data alone it is faster to send two 127-octet messages than one 128-octet message (or whatever the critical size is). The data show a mean dispersion, but high reliability. Only a single packet was lost in over 1100 volleys. I conclude this a good data set to experiment with for TCP tuning. Earlier I FTPed three megabytes of useful data to the Stuttgart fuzzball using the battered and scarred fuzzball TCP, but with reasonably good results (I didn't have the chance to capture retransmission counts, etc., with that transfer). Therefore, I would class this data set as a good example of a network design that tries to maximize reliability at some cost in delay. The tests were done on Sunday afternoon. I'll do another one on a day when things are falling apart. Both the raw data and Sun-format scatter diagram are on udel2.udel.edu in anonymous/guest as the files usecom.txt and usecom.bit respectively. Dave -------