Re: Comments on LARIAT and Comcast not same problem

[David Farber <dave () farber net>]

________________________________________
From: Tony Lauck [tlauck () madriver com]
Sent: Tuesday, February 19, 2008 11:29 AM
To: David Farber
Subject: Re: [IP] Re:   Comments on  LARIAT and Comcast not same problem

Van Jacobsen's mechanisms are finely tuned and subtle. In no sense are
they a kludge. There are a few things that they don't do, admittedly:

1.  They can not deal optimally with poorly architected, implemented or
operated data links that have significant data link error rates.  These
are problems that need to be corrected in layer 2.

2.  They can have no concept of "user" and hence no concept of per user
fairness, given that "user" is not something in the TCP/IP architecture.
In particular the relationship between an IP address/port number and a
user can be complicated because of mechanisms such as NAT and a variety
of possible business arrangements that subscribers have with their
suppliers. Multiple connection bit torrent applications are just one of
many complications in the mapping between TCP connections and "users".

However, at least with certain data links (including all point to point
links of which DSL is an example), if all traffic from one "user" goes
through a single router, that router should have sufficient information
to implement per user fairness policies, by invoking appropriately
selective packet drop mechanisms, which will then work with Van
Jacobson's TCP mechanisms. While explicit congestion notification,
particularly multi-bit mechanisms that provide rate estimates, may
respond quicker to rapidly changing traffic requirements than packet
drop based mechanisms, they are not necessary to deal with bulk traffic,
such as bit torrent transfers which take a very large number of round
trip times to complete.

The DOCSIS 2.0 data link appears to have insufficient capabilities to do
per user fair traffic allocation on the uplink. This is too bad, because
the protocol was designed at a late enough date and has sufficient
complexity that this could and should have been done. Hopefully, version
3.0 will correct this.


Incidentally, DOCSIS is not the only data link that would have problems.
  The original Ethernet CSMA/CD would probably have similar difficulties
today, if not worse, if we were still running at low speeds, half
duplex, with traditional CSMA/CD.



David Farber wrote:
> ________________________________________
> From: Richard Bennett [richard () bennett com]
> Sent: Tuesday, February 19, 2008 5:51 AM
> To: David Farber
> Cc: ip
> Subject: Re: [IP] Comments on  LARIAT and Comcast not same problem
>
> It's naive to assert that Jacobsen's mechanisms have served us for 20 years since they've actually been revised at 
> least twenty times and still fail to address the fundamental problem of per-user fairness. The most recent revisions 
> address a perverse behavior when one or more links run WiFi, where packet loss is typically the result of a high 
> intrinsic error rate rather than congestion. The WiFi variation ignores the first packet drop from the 
> decrease/increase effect on window size. Jacoben's fundamental error is overloading packet loss with congestion 
> signaling instead of adding an explicit congestion signal into the IP layer.
>
> The problem that network operators have today is to prevent small numbers of users from using disproportionate 
> bandwidth, which isn't really the same as protecting the Internet's interior links from congestion. This problem was 
> unsolved by the TCP/IP architecture before and after Jacobsen's kludge.
>
> RB

-------------------------------------------
Archives: http://www.listbox.com/member/archive/247/=now
RSS Feed: http://www.listbox.com/member/archive/rss/247/
Powered by Listbox: http://www.listbox.com