benchmarking and performance tuning
Thu Jul 23 12:58:10 1998
Thanks for all the feedback.
I tried an experiment which yielded an interesting result:
If I set the window size to 6000 (4xMTU) the performance rockets all the
way from 19 Mbps to 39 Mbps. A pretty substantial improvement. Now for
the bad news, this happens on transfers to /dev/null only. On transfers
to a partition the bigger window size actually makes things worse!
This seems to suggest several things:
1. That there are optimal values of mtu, segment size, and window size
for any configuration. They may not be readily derived, but some
tweaks are possible.
2. One of my machines is simply too slow. The fast one is running a K6-300
at 4x83 with 41.5 PCI clock whereas the slow one is merely a P5-133 at
2x66 with 33 PCI clock.
I'm going to upgrade the slow system to match the configuration of the
fast in the near future and this should help. If anyone knows of a technique
for optimizing the parameters of item 1. please let me know.
BTW, where is ttcp.c to be found? I've seen it mentioned in several places,
but without links.