Outbound Route Optimization
Mike Lloyd
drmike at routescience.com
Tue Jan 27 13:57:03 UTC 2004
Michael.Dillon at radianz.com wrote:
> You hit the nail on the head. Fundametally, any route optimization
> technique that tries to treat an aggregate of the network as
> a blob which can be measured will suffer the same type of problems
> as an IP over ATM network.
Observationally, IP over ATM has interesting latency effects. But it's
rather defeatist to tar "any route optimization technique" with this
critical brush. First, real products exist which do not require
treating the /8 as a contiguous block - local (no-export) more specific
injection is well understood (for those who believe it's even
necessary). Second, to the extent that your point is about variance of
one quantity of interest (latency), there are comparatively ancient
statistical techniques for handling that - hardly new science.
> There will always be a hidden layer
> of complexity that will affect your traffic flow and which you
> cannot influence.
Surely the same statement is true of almost all other products we can
buy, not just bandwidth? There's hidden, important complexity in how
airlines operate, how coffee is produced, how cars function. Somehow,
though, we manage to make price-performance tradeoffs about these
commodities all the time.
In a network of networks, we cannot perform "strong TE" across everyone
else's backbone. It does not follow that it's pointless to measure the
other networks we rely on.
> The same thing will happen with any measurement system that tries
> to classify a path through someone else's AS. You have no control over
> what happens in that AS and, more importantly, you have no control
> over the peering points bewteen ASes. Your measurements are as
> meaningless as measurements of an IP over ATM network.
So I agree with you that a /8, measured together, has a high signal to
noise ratio. You're right that giving an extremely tight latency bound
on a /8 aggregate is often impractical. It does not follow that it is
impossible to extract signal from the noise; specifically, we routinely
find concrete and stable (hours to days) instances where aggregate
packet loss for a /8 will vary between various paths to reach it.
No doubt someone hereabouts will scream "so buy more access", as if that
means you can stop looking at off-net price/performance. For anyone who
believes they don't need to look, fine - enjoy. For anyone who wants to
evaluate the bandwidth they buy on an ongoing basis, tools exist,
ranging from traditional network management pieces through to route
optimization. The question, surely, is not whether off-net measurement
is possible; it's whether it's useful to automate feedback from such a
system into the network.
Mike Lloyd
CTO, RouteScience
More information about the NANOG
mailing list