[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Deadlock free routing

In message <33B26B1F.53278A0C@xxxxxxxxxx>, Richard Beton
<rdb@xxxxxxxxxx> writes
>One of the exciting corollories of these theoretical results is the
>rather counter-intuitive discovery that <I>less</I> buffering is needed.

Depends on your starting point. 

There is a close analogy between these things and manufacturing. 50 years
ago, everyone thought that you needed to maximise throughput on a
production line and to hell with latency, quality, change-over time, etc. To
meet this throughput they used large buffer stocks, and the factories were
like warehouses that occasionally processed the stock.

Now, manufacturing is designed to maximise quality, minimise delay, and
particularly to optimise use the working capital by turning over the inventory
many times per year. The way they do this is with "Just In Time", minimal
batch sizes, and minimal buffers.

With links, Neil Davies at PACT has found that the immediacy of the flow-
control of 1355 links means that as soon as a path is blocked, an alternative
path can be chosen; if there was extra buffering, it would increase the control
loop delay, and that reduces the responsiveness of the system. Of course
you need some buffering, but the most importnat thing in switched networks
is the valency of the switch, and this is maximised by minimising the amount
of logic and buffering for each port of the switch. 

Many of us will, of course, always argue for more buffering, as I was in a
discussion earlier this afternoon!  But there are precedents for things being
better with less buffering.

Paul Walker                      4Links                      phone/fax
paul@xxxxxxxxxxxxxxxxxx          P O Box 816, Two Mile Ash    +44 1908
http://www.walker.demon.co.uk    Milton Keynes MK8 8NS, UK      566253