too long for a subject, but more descriptive:
receiver-throttled tcp, and
tcp on large bdp (bandwidth delay product) networks
i submitted a patch (tcpnewreno) that deals with two seperate
problems i've seen when streaming over tcp. (and one discovered
while debugging.) the tool used for debugging (nettest(8)) has
also been submitted.
cavet: for rpc-like protocols (e.g. 9p) that don't use multiple
oustanding (e.g. the mnt driver), these changes will likely not
have a large effect. this is going to be seen mostly in streaming
senerios.
the problems:
0. the current tcp stack loves to send less-than mss-sized segments
and can send tinygrams for no apparent reason.
1. if the reader on the receiving end is reading so slowly that the
receive queue fills, the connection can slow or even livelock.
1/10th to 1/50th calculated bandwith can be seen.
2. much lower than expected (again 1/10th to 1/50th) throughput
can be seen on large bandwidth-delay product networks.
the solutions:
0. nix had a fix for the tiny gram problem. i'm not sure who did
the work, but thanks. good stuff.
1. corrections to zero-window probing fixed the receiver-pacing
and livelock problem. (some of this was included in nix.)
2. NewReno was fully implemented and a few rfc-compliance
issues were addressed. with the constants in the submitted code,
BDPs of up to 1mb should be doable. that should be enough for
100Mbps @ 100ms. i'd be curious to know if anyone has a longer
or fatter network. you can use the mathis* formula to calculate the
theoretical speed of your network as
bw = √(3/4)*(rtt/mss)* (1/√p)
where
bw bandwith
rtt round-trip time
p probability of loss event (i.e. entering newreno
recovery, not the probability of a lost packet)
obviously this doesn't work for p=0, and gets
questionable for p<ε, for some ε = f(bdp).
("the macroscopic behavior of the tcp congestion avoidance algorithm",
Matthew Mathis, Jeffrey Semke, Jamshid Mahdavi,
ACM SIGCOMM, volume 27, number 3, July 1997. ISSN # 0146-4833.)
a detailed discussion of the NewReno changes and observed performance
is here
http://www.quanstro.net/plan9
- erik
