[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
RE: A Hardware Question
The obvious question to ask is: how much raw processing power do you need?
Will your thousands of processes each need several/many MIPS of CPU, or are
they (individually) relatively low-powered.
I ask because as Peter Welch has shown, KRoC can support many (100,000+)
processes very well on modern CPUs, and certainly used to support SMP
setups. The overheads associated with KRoC are much lower than "normal"
threading setups. There are also options to use Ethernet should that be
But as you'd guess, this only helps if the computational power of, say, a
modern 4-way Opteron cluster (or, cheaper, a dual core Athlon-64) is
sufficient. I note Eric's suggestion of a DSP, and Transtech might indeed
have useful experience to relate. But he's right in considering them harder
to program well.
> Okay folks, I have a question about buying parallel hardware.
> If I were wanting to buy hardware today, and I wanted fine-grained
> (-ish) parallelism, what would I do? (Oh, btw I am asking you folks
> because I really would want to run occam on it).
> My problem is that I don't think that an OTS cluster of workstations
> linked by fast ethernet would be good enough, as you might able to see
> from the following specs:
> Processes that were numerous (typically 1,000 to 10,000) but fairly
> compact (say a thousand bytes of code each, perhaps more).
> Communications that are small packet-size, with frequent exchanges (does
> me no good to save 'em up and send a book-sized transfer, I need lots of
> little packets going back and forth).
> Not much locality, so any one of my processes can want to talk to any
> I am focussed on high-level issues these days, so I am out of touch with
> the hardware that is available. Does anyone know if there is stuff out
> there that would be appropriate?
> I should tell you that my preference is to use Macintosh hardware, for
> programming and interface reasons, but I realise that that might keep me
> separated from occam for a long, long time.... :-(
> Any thoughts, suggestions or hoots of laughter would be accepted
> Richard Loosemore.