rbriscoe at jungle.bt.co.uk
Wed Aug 8 19:44:41 BST 2007
At 23:45 07/08/2007, dimitri papadimitriou wrote:
>>- Is a good system & equipment design goal to sacrificially throttle
>>bit-rate to avoid CPU overload (so the brain can still work out what to
>>do about congestion)?
>what do you mean with "sacrificially throttle" ?
I guess I didn't explain myself well. We'd like researchers not to have to
think up alternaitves to smart congestion control (e.g. XCP or even AQM)
that achieve the same effect, but can be implemented in cheap, lower layer
equipment. The research issue is whether we will be able to rely on the
same 'sacrifical throttling' as we do now to avoid this issue.
Example 1/ A route lookup engine requires a certain min amount of time to
do the most complex address lookup. Usually the line feeding a route lookup
processor is sized so that even a stream of the smallest possible packets,
with the lowest possible cache hit rate, couldn't arrive at the route
lookup engine faster than it could look-up the addresses. In this case, the
bit rate of the line is sized to effectively sacrifically throwing away
excess packets to protect the route lookup processor.
Example 2/ Even if all the L3 routers feeding into the L2 network were
outputting at line rate and the traffic matrix was the worst possible - all
converging on the lowest capacity L2 switch - the capacities could be
arranged so that any L2 switch could not ever be overloaded (aside from
failures). Then if the operator wants everything to do AQM or ECN or
something, only the L3 routers need to do AQM. The L2 switches don't have
to have any AQM code, because they will 'never' get overloaded. So the L3
routers are effectively sacrifically throttling traffic on behalf of the L2
Example 3/ A single transmission line is actually a degenerate case of both
the above examples, where the queue feeding the line sacrificially
throttles traffic so that the line doesn't have to discard excess traffic
at the phy layer (where the analogue signal would screw up, instead of the
queue discarding whole packets). However, this is an example where a
processor is protecting a phys component, rather than the other way round.
My point was simply that we should be able to rely on sensible system
design and not have to say that deploying smarts at all layers is a
research issue. But, for the resons below (about packet sizes), we might
have to worry about how to put smarts in lower layers as well.
>>- distinguishing per-packet congestion from per-byte congestion (similar
>>problem to distinguishing congestion from wireless transmission losses)?
>you made that comment during the "corruption loss" discussion, do you see
>this as a separate discussion point now ? note also relationship with the
>packet size (former).
It's a separate discussion point. It just has strong parallels with the
corruption loss issue (ie a solution to one might solve the other).
>>- ever-widening range of packet sizes: if design equipment for larger ave
>>pkt size, gets overloaded with flurries of small packets.
>you made that comment during the "small packet size" discussion -> i
>guess this would like to see a discussion about impact wrt relative avg
>size and range ?
Yup. Ideally as links get faster, we'd like to only have to deal with
larger packets, but we're never going to be able to have a flag day
stopping people sending smaller packets (partly because there will always
be slower links connected to faster links in the Internet).
This is likely to become a big issue in the future when considered against
the first point about sacrifical protection of processors by bit rate
throttles: packet header processors will always have to be able to handle
streams of packets with 0byte payloads.
Currently, we tend to assume congestion is by bits. But these worrying
thoughts imply we could start to see more per-packet congestion appearing
alongside per-bit congestion. That complicates matters greatly.
Background to this issue is discussed in
draft-briscoe-tsvwg-byte-packet-marking-00.txt that I recently posted. But
that I-D only hints at the 'sky is falling' thoughts of this email.
Bob Briscoe, <bob.briscoe at bt.com> Networks Research Centre, BT Research
B54/77 Adastral Park,Martlesham Heath,Ipswich,IP5 3RE,UK. +44 1473 645196
More information about the Iccrg