how-to block ads
|
DSLrgm
Premium,MVM
join:2002-08-22
Oak Park, MI
| TCP Koolaid
From a colleague of mine:
The fundamental problem with each of these, for the web and email, is that they don't necessarily play well with TCP - you smash the installed base. A good paper for their comparison is »www.mit.edu/~chuvpilo/papers/chu···ject.pdf. His comment is amusing: if you have a network that is 10% loaded, XCP and HighSpeed TCP are completely benign, because TCP still gets about 10% of the network. Yes, but what if you didn't have that level of overload? TCP might well still get 10% of the network...
You can also look at:
FAST: »netlab.caltech.edu/FAST/
QuickStart: A. Jain and S. Floyd. Quick-Start for TCP and IP, August 2002.
IETF Internet draft »www.ietf.org/internet-drafts/dra···t-02.txt & »www.icir.org/floyd/quickstart.html
HighSpeed TCP: S. Floyd. HighSpeed TCP for Large Congestion Windows, August 2002.
IETF Internet draft
»www.icir.org/floyd/papers/draft-···-00c.txt & »www.icir.org/floyd/hstcp.html | |
heatfan
join:2001-01-10
Fort Lauderdale, FL
clubs:  |  Thanks for the links to the additional information, these links look very interesting ... | |
fgoldstein
join:2003-01-21
Newton Highlands, MA
 ·RCN CABLE
| reply to DSLrgm
This is interesting stuff, but especially because it took them so long! I do recommend the chuvpilo paper; it is only 8 pages and quite readable.
We were doing this type of work back at DEC in the 1980s. DECnet Phase V, defined by 1986, had explicit congestion notification, which was used to tweak the TCP-equivalent protocol (TP4). I managed to get it added as an option in Frame Relay (the FECN bit), but it saw little use, especially because TCP had no way to use it, since IP didn't pass it through. Now we're finally seeing the IETF crowd running simulations very similar to the ones we ran at DEC about 15 years ago! And waddayaknow, they get similar results.
QuickStart, which Amit Jain and Sally Floyd are now proposing, is very similar to Fast Burst Reservation, which was first proposed in the ATM world in 1990 and then again around 1992. A good idea, if I can say so as the first inventor, but it is a little harder to do in a connectionless network like TCP/IP than in ATM. Again, old DEC ideas pop up long after being pooh-poohed as NIH.
Frankly the whold TCP/IP protocol stack is creaky and in need of replacement or major overhaul, but the inertia against that is huge. | |
|
