On 11/30/06, Ian McDonald <ian.mcdonald@xxxxxxxxxxx> wrote:
Folks, CCID3 is fucked up. Sorry to put it this way but thats how I feel. With 2.6.19rc6 plus the patches from http://wand.net.nz/~iam4/dccp/patches/ it works fine. (The only significant patch here is tx_qlen - all the rest should have no impact). So here are the stats using iperf: Test 2.6.19rc6+ian 2.6.20-acme20 Direct connect 87 Mbits 36 Mbits Through a Linux router (1.2 ms rtt) 36 Mbit 200 kbits (yes that is correct!) Through netem (100 ms rtt, 5% loss) 130 kbits 116 kbits (OK) So somewhere on low rtt links we've stuffed up. Below is the logs from the receiver from when performance dropped from 36 Mbit to 200 kbits. The illegal ACKs are documented in another email and I don't think are the problem. The issue I suspect is that it says rtt is around 12000 usecs which is 12 msec and it is really 1.2 msecs.... I can't look at this for a while as have to go to work and have my kids over this weekend. I'll start bisecting next week if I can if nobody else has fixed. Can I suggest that people test CCID3 on a range of loss, rtt combinations please? CCID3 has proven time after time to be very sensitive....
As I told you, I'm preparing a more comprehensive test environment here, just being sidetracked by some problems using net-2.6.20 on a parisc64 machine. - Arnaldo - To unsubscribe from this list: send the line "unsubscribe dccp" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
