Mark Allman: Internet measurement: what next?

Tue Jul 8 15:23:23 UTC 2003

Eddy,

E.B. Dreger wrote:
> You mean something like RouteScience [supposedly] does or ECN
> influencing next hop?

Guilty as charged, of course :-)

> Yes, I know, that would get ugly for transit ASes.  

Well, the measurement itself isn't ugly - it's just a question of what 
you choose to do with the information.  Some more conservative folks 
just take the data over time, and use it for peer or transit 
elimination, for example.

> For the degenerate case of a leaf AS, though,
> things become simpler...

Degenerate seems a little strong, even in its mathematical sense - 
rather like an auto maker calling car drivers the "degenerate" case of 
mechanics.  But on this list, I suppose I'll let it pass :-)

I'll agree that the constraints on active use of measurements for route 
change are lighter when you have no BGP downstreams.  But it's still 
possible to use good end to end measurements, with very strong damping, 
to make changes in a transit routing context.

Others in this thread have been talking about bandwidth measurement 
along the variety of paths you use across other peoples' networks.  To 
me, that's a good answer to one of Mark's original questions:
   What do we need to be able to measure that we cannot measure
     very well today?
That is, I consider end to end bandwidth one of those quantities we'd 
love to measure widely and often, but practically, we cannot.  Available 
techniques are ugly and unscalable, but more disconcertingly, estimates 
of bandwidth appear to have a shelf-life comparable to certain 
sub-atomic particles.

That's not to say it's a lost cause, of course.  Many end to end paths 
are constrained primarily by first hop (which you can usually measure 
directly with SNMP) and last hop (which is roughly constant, and next to 
impossible to influence).  For the remainder of the path in between, 
it's important to characterize it, but accurately assessing bandwidth is 
hard.  Measuring probability of packet loss over time is a very 
practical proxy.  It won't give you all that precise a projection of 
your goodput if you throw, say, a multi-gigabyte ftp down the pipe, but 
for short or light transactions, it's not bad at all.  After all, once 
you've accounted for first and last hop, your traffic is probably a 
pretty minor fraction of total load on each link along the chain.

Mike Lloyd
CTO, RouteScience