richb.hanover at gmail.com
Tue Apr 30 22:39:43 EDT 2013
I noted a paradoxical effect using the blip tool and CeroWrt. I wrote it up and included screen shots in: http://www.bufferbloat.net/issues/430 Here's the text of the note.
Experimental setup #1
- Macbook connected to demo3.intermapper.com, where I executed iperf -s
- Macbook running iperf locally to that server (iperf -c demo3.intermapper.com -i 1 -t 30)
- Macbook running http://apenwarr.ca/log/ connected to primary (bloated) router running DD-WRT from a couple years ago
Result: blip chart times increase during iperf as expected. The response goes from ~120/200 (green/blue, respectively) to ~500msec for both. (see "Test #1 using primary router wireless" image)
Experimental setup #2:
- MacBook is connected to CeroWrt 3.8.8-4 which is connected to LAN side of DD-WRT which then goes to the Internet.
- Run "iperf -c" as before
Result: the blip chart times seem to exhibit approximately the same behavior, with the same high response times during the iperf transfer. (See "Test #2 using CeroWrt5 wireless" image below)
Both charts seem similar - this is counter to my intuition, and counter to my experience using a simple sequence of one-second pings which don't vary much from the baseline in the experimental setup #2.
- blip is a cool hack, but a bad test, and doesn't accurately represent the responsiveness of the network.
- fq_codel as implemented in CeroWrt3.8.8-4 does not handle these "TCP Pings" well/as expected.
More information about the Codel