HTTP server choices?
cdshaffer at acm.org
Thu Jan 5 03:34:32 UTC 2006
Alan Knight wrote:
>I know you're not paying attention to settings, but I feel I should note that the VW server is set to progressively throttle connections with a default upper bound of 300 (on the assumption that it is better to serve stably than to try and handle as much as possible). So an upper bound of 270 sounds like it could be a throttling artifact. And I believe VW's file I/O is also VM blocking, so it doesn't seem like that has to be the major obstacle.
You are certainly correct in that I don't want to explore all of the
dependencies of each server's performance on its various configuration
parameters. In the case of my more recent work I did explore both
adjusting the maximum # of simultaneous connections (I experiement with
300, 400 and 500/s) and the maximum connection lifetime (100ms, 200ms
500ms) but not all combinations of these. The results deviated very
little. One thing that I did notice about VW/Webtoolkit recently is
that its throughput is better (time from the first TCP packet to the
last is very short) and that it survives much more stress than the
previous versions did.
If you want my data and graphs just let me know. I am also happy to
provide my scripts and test image. The format is different than my
previous benchmarks but the files are reasonable self-describing since
they are still based on httperf output.
More information about the Squeak-dev