Work with Karsten on TorPerf Experiments
This ticket is to catalog my portion of the work with karsten on #2769 (closed). I plan to help with a few experiments:
- Determine new guard cutoff: 2pts
- Determine optimal CircWindow: 6pts
- Throttle clients network-wide: 4pts
- Disable EWMA: 1pt
- Extra luck point: 1pt
Activity
For the throttling experiment, I think we want to run at least 2 experiments, but probably more.
We want to be able to support streaming video on youtube, which for 360p HTML5 video takes about 35-40K/sec, or about 20M for a 10min video.
I think this means we should try to see what a 10M/40K burst/rate looks like in torperf, as well as a 100M/10K limit looks like.
The idea is we want to test to see if minimally invasive limits can still choke out the jerks and improve performance.
-
Here is the experiment log for the bwconn stuff done so far:
perconnbwrate=40960 perconnbwburst=10485760 bwconn start: Fri Apr 22 09:57:34 UTC 2011 bwconn end: Sun Apr 24 01:57:57 UTC 2011
perconnbwrate=40960 perconnbwburst=104857600 bwconn start: Tue Apr 26 11:32:40 UTC 2011 bwconn end: Fri Apr 29 05:26:20 UTC 2011
perconnbwrate=20480 perconnbwburst=104857600 bwconn start: Fri Apr 29 05:26:20 UTC 2011 bwconn end: Sun May 1 21:23:18 UTC 2011
We need to keep a wiki for this info..
This work is also not done, but this ticket was just for accounting my part of it..
Trac:
Points: 14 to 4
Resolution: N/A to fixed
Status: new to closed