[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: rewriting CSP processes
----- Original Message -----
From: "B.M. Cook" <b.m.cook@xxxxxxxxxxxxxx>
To: "Lawrence Dickson" <tjoccam@xxxxxxxxxxxxx>
Cc: "B.M. Cook" <b.m.cook@xxxxxxxxxxxxxx>; <IanPage99@xxxxxxx>;
<java-threads@xxxxxxxxx>; <occam-com@xxxxxxxxx>; <ian.page@xxxxxxxxxxxx>
Sent: Friday, 29 September, 2000 1:26 AM
Subject: Re: rewriting CSP processes
> > In all this discussion of multiple clocks and
> > unclocked logic, I think this is the first mention
> > of metastability.
> > That has always worried me, in all computer
> > design including occam and CSP. Transforming a
> > smooth (time) function into a step function, and
> > deciding any race... Is it true there is always
> > a non-vanishing chance of failure?
> In a fully synchronous (i.e. all functions driven by the same clock
> and all logic delays small enough) system there is no problem.
> In an asynchronous system - very common - there is a metastability
> It is statistical and leads to a probability of failure. It is now well
> understood and we know what measurements to take to predict the
> of failure. Some of the more recent chips have an inherently low failure
> probability. A widely used technique to re-synchronize signals is the
> multi-stage latch where each additional stage reduces the probability of
> error by about 4 orders of magnitude from the inherent already low
> probability. You can choose the number of stages to achieve your target
> error rate, in practice 2 stages is usually enough - that's just one extra
> latch per signal. The extra delay does add latency and hence reduces
> throughput. There is no problem letting a synthesis tool put the logic in
> for you.
> Yes, it is true that there is always a non-vanishing chance of failure,
> it can be made as small as you like.
There's also a finite probability of simple logic levels being "wrong" due
to noise margins etc.. in the semiconductors which practically sets an
end-stop on single device reliability. ISTR something of the order 1:10^14
for TTL but it's a while since I've worked with the numbers...........
-- Stephen Maudsley, Esgem Limited mailto:Stephen.Maudsley@xxxxxxxxx
-- Tel: +44-1453-521626 Mobile: +44-7770-810991
-- company registered in England 3372135, http://www.esgem.com/contact.htm