zweig@cs.uiuc.edu (Johnny Zweig) (04/12/91)
I was just thinking about running benchmarks/test-suites for TCP/IP and realized that I haven't noticed any really thorough traffic characterizations in the literature. I have in mind things like a probability-density function for number of octets sent during a connection, possibly correlated with some burstiness measures and mean throughput. Then there are even cooler things to wonder about, for instance, what percentage of connections are to passive (listening) sockets (I would guess 99+%)? And what percentage of connections follow the "A calls B, B finishes and closes its end of the connection, A closes in response" pattern? How often do SYNs and/or FINs "cross" in the network? What percentage of TCP packets ever require retransmission? I could go on, but you get the point. If I wanted to exercise my TCP/IP in a "realistic" way, I would have to argue that my load-generator did "screwy" things (like active connect-requests crossing in the mail) more or less the right percentage of the time (or enough not to throw all my numbers off). I could see things such as the average number of TCP connections mattering to a benchmark (a naive search algorithm for finding connection descriptors only bites you if there are more than a few of them active at any time, for instance). Does anyone know of what sorts of traffic characterizations for TCP/IP there are out there? Just wondering... -Johnny Curious