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

Re: Transputer Development System, 2006? (aka Sony PS/3 runs multi-core Linux)

tjoccam@xxxxxxxxxxx wrote:
Barry wrote:
I'm still trying to find the compelling reason that will convince the
to jump to parallel software

I've got to try to do this for DARPA... My thoughts thus far:

Transparency (not hiding the most important part of the design, like
interrupts and peripherals, in "drivers" just because they involve real
parallel hardware and function);

Real components (things that you can hook together and get genuinely
predictable behavior that is the join of the behavior of the parts -- I do
NOT mean the things OO laughably calls "components", but real pieces as
found in an automobile);

As a consequence, this kind of programming does not hit a complexity WALL.
If done wrong, of course, it worsens the complexity wall -- which is why
it has the reputation it has.

Larry Dickson
It's the way the "real-world" works that's always makes process notations essential for me.

I think over the last 100 years we have accepted that the Universe has a notion of locality and that an observer at one point in space-time only observes what's happening at another by passing messages (I'll quietly ignore quantum entanglement until someone tells me they have computing system communications using it). As soon as your system is distributed enough that you're forced to consider locality then you are using processes whether you realise it or not. And if your system only behaves like a process then you'd better use a notation that captures that.

I guess that strictly you're using processes when you have a system with more that one clock domain (all Einstein's stuff was about crossing moving clock domains). For physically small systems where the transit time of messages is small enough to be considered zero (or fully synchronous perhaps) then you can get away without thinking about processes. With processors tending to nS instruction times, I guess the maximum distance between system components (including the "rest of the universe") is of the order of 15cm.

So there you have it - Darpa can only NOT use processes in their system designs if their whole universe is less than 15cm - how big is the United States?

I could probably make this less rambling with more coffee...