[e2e] It's all my fault
lachlan.andrew at gmail.com
Tue May 15 16:27:19 PDT 2007
On 15/05/07, Vadim Antonov <avg at kotovnik.com> wrote:
On Tue, 15 May 2007, Lachlan Andrew wrote:
> > selecting a small number of parallel paths, and regularly reselecting
> > based on performance (a la BitTorrent), is as good as selecting the
> > best path.
> Mmmm... and that solves the problem of an end host not having an idea of
> what backbone topology is like - how?
> This is all neat when you enter those points manually, but try to sell
> that to the actual users.
I agree that the route selection should be automatic, but it doesn't
need to be done by the network.
I'm advocating very loose source routing here. The key to this is
diversity. That diversity can come from knowing one host on each of
several nearby ISPs, to "force" your traffic onto that AS. My
understanding is that different ISPs will usually have quite different
paths to the destination, primarily through their own networks (but
I'll gladly be corrected).
Given three stepping stones, which depend on your own location rather
than that of your peer, you get three almost independent routes, which
gives you p^3 chance of hitting temporary congestion, where p is
the chance of any one link being congested.
The "local stepping stones" could be signalled to individuals by DHCP
the way DNS servers are, or be provided by a database like
<http://www.geobytes.com/IpLocator.htm> if ISPs don't want to
advertise the competition.
Also remember that most traffic is going from some infrastructure to a
user (even if that infrastructure is a P2P session or some such). The
efficiency gain comes if that infrastructure (not the user) has some
control over its route, if it chooses to.
My understanding was that this thread was about whether or not the
standards *allow* source routing. Most home users will not want to
bother, and should not be forced to. However, if Don Towsley is right
that TE can be greatly simplified by a handful of big traffic
generators doing their own load balancing, why should we take that
existing functionality out of the standards?
There is clearly a case for not *forcing* equipment vendors to put it
in hardware, or carriers to enable it, the way existing IP options are
often treated, but it seems useful for IPv6 to say "if you want to
implement source routing, this is how it is done". If it is there, it
can be tamed (say by limiting the number of hops that can be
specified), but if it is not in the standard at all, content/service
providers who have legitimate uses for it can't use it even if they
choose to. Currently, I believe Akamai uses its servers as routers to
implement its home-brew source routing; wouldn't it be better if
companies could use routers as routers?
Lachlan Andrew Dept of Computer Science, Caltech
1200 E California Blvd, Mail Code 256-80, Pasadena CA 91125, USA
Phone: +1 (626) 395-8820 Fax: +1 (626) 568-3603
More information about the end2end-interest