home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!stanford.edu!ames!haven.umd.edu!mimsy!partho
- From: partho@cs.umd.edu (partho pratim mishra)
- Newsgroups: comp.dcom.cell-relay
- Subject: Hop by hop vs End to end flow control
- Message-ID: <62141@mimsy.umd.edu>
- Date: 17 Nov 92 15:35:32 GMT
- Sender: news@mimsy.umd.edu
- Organization: U of Maryland, Dept. of Computer Science, Coll. Pk., MD 20742
- Lines: 59
-
- craig@sics.se (Craig Partridge) writes
-
-
- >>"Allen Robel" <cell@mythos.ucs.indiana.edu> writes:
-
- > they compared
- > their hop by hop (HBH) scheme of flow control to TCP as implemented in
- > 4.3-Tahoe BSD (they also compare HBH with the modified window adjustment
- > in the 4.3 BSD Reno version of TCP). Their results seem to indicate
- > that, in a network with a high bandwidth-delay product, TCP exhibits
- > very large oscillations in end-to-end delay, and takes about 100
- > round trip times (losing about 190 packets) before a link becomes
- > fully utilized. In contrast, HBH loses NO packets and maintains
- > a very even end-to-end delay variance. Note that I'm not condoning
- > their scheme. I'm just saying that there are schemes out there that
- > do better in the area of flow control than current implementations of
- > TCP in high speed networks.
-
- >> I'd just like to note that the jury is still out on these results.
- >>In general, the literature (both theory, simulation and practice) shows
- >>end-to-end flow control outperforming hop-by-hop. And I've heard folks
- >>question whether the representation of TCP-IP in the model was accurate.
- >>Certainly, given the weight of results one way, it is not clear that one
- >>dissenting paper should be sufficient to lead us.
-
- and later Rob Warnock writes
-
- >The oscillation is real, and quite large; I have seen it during FTPs with
- >"hash" turned on. But also note that Jon Crowcroft (et al?) published a paper
- >some time ago which proposed a simple change to TCP slow-start that smooths
- >out most of the oscillation. I wouldn't count TCP (or end-to-end congestion
- >control) out yet...
-
-
- There seems to be some misunderstanding about the "accuracy of the
- TCP model". We used the 4.3 BSD Tahoe version of TCP as our model.
- In fact the simulator we used was written by Lixia Zhang and used in several
- previous studies of TCP (which appeared in CCR and Sigcomm). What
- we found was there were many reasons why 4.3 BSD TCP would perform
- badly over high speed links (45 Mb/s and faster). For lack of space
- we had to leave out most of the explanation in our Sigcomm paper,
- although a companion Technical Memorandum examines the issues in some
- more detail. 4.3 BSD Reno and selective acks togther do solve some of the
- problems though not all. I would be glad to discuss with interested people
- what it does and what it does not fix.
-
- As far as the hop by hop vs end to end argument goes I believe we are
- entering the domain of religion so I will hold my horses. However, I
- would like to disagree with the claim that "the literature (both theory,
- simulation and practice) shows end-to-end flow control outperforming
- hop-by-hop". One has to compare specfic mechanisms before making such
- a sweeping statement.
-
- -Partho Mishra
- --
- | Systems Design and Analysis Group
- Partho Pratim Mishra | Department of Computer Science
- partho@cs.umd.edu | University of Maryland
- | College Park, MD 20742
-