AW: [e2e] Queue size of routers

[Vadim Antonov]

On Fri, 17 Jan 2003, Michael Welzl wrote:

> > > Routers in real backbones have the delay*bw of buffer space.
> > 
> > good!
> 
> Why?
> 
> I'm serious - I know that a delay*bw queue length is just
> right if, for example, you suddenly fill the capacity of a
> dumbbell bottleneck in a simulation with new flows and
> don't want some of the initial packets to be dropped,
> thereby eliminating a potential traffic phase effect. But
> is that a good choice for a backbone router?
> The LA<->Tokyo RTT is quite a bit of delay ... is that
> really reasonable when only a small number of flows may
> show this RTT? Shouldn't backbone routers be more concerned
> with traffic aggregates instead?
 
Well, what you have for "delay" is characteristic RTT of the traffic mix,
plus some extra to accomodate deviation; not the maximal RTT. I'm not
aware of any tool for picking that parameter, it's pretty much "gut
feelings" of backbone engineers.

My impression that queue size is often set too high, causing unnecessarily
high latencies in case of congestion.

[Historical note -- the early cisco 7000s and nearly all ATM switches had
insufficient buffer space... which led to really horrible performance in
practice. Flow smoothing may be present, but it merely reduces probability
of congestion at a given average load; when circuit gets congested you
still need RTT worth of buffers to accomodate delay in the end-to-end
control loop.]

> And: is there RED in backbone routers?

Yes. Both Cisco and Juniper have weighted RED.

RED pre-empts congestions by initiating reaction earlier, that allows to
shrink the queue size, and thus make path latency variation smaller (thus,
in turn, reducing TCP retransmit timeouts and time needed to recover from
congestion).  RED also reduces effects of flooding attacks.

--vadim