Re: [Nomen-dev] Multithreading
Brought to you by:
bhurt
|
From: Brian H. <bh...@sp...> - 2002-04-09 22:48:16
|
Dave: this started life as a programming language for the Thinking Machine. Which, actually, makes it precursor to what I think is going to be the future of computing. The current generation of CPUs is hitting a wall for ILP in a single thread, so they're going to Symetric MultiThreading (Hyper Threading is you're an Intel marketroid) to increase the ILP yet farther. So we're getting more threads per CPU. And ccNUMA is becoming defacto for multiprocessors, even on the low end (HyperTransport is ccNUMA, I don't care what AMD's marketroids say. It's just not very non-uniform). On the high end, they're using it as an excuse to push the number of CPUs yet higher. I would not be at all surprised that within 5 years Sun is selling a box of 128 CPUs, each with 4-way SMT. And that 5 years after that, they're selling a box with 512 CPUs each with 16-way SMT. That's 8,192 processors- basically, we're back to the Thinking Machine. That being said, I think cilk's main fault is that it's based on C, not some sane language. Of course, Java didn't exist back then, and basing it off anything else would have meant it would have been marginalized even harder than it has already. But I think that combining a virtual machine with the idea of ultra-light-weight threads has a lot of merit. I have some reading up to do. I have reading and thinking to do. Brian On Tue, 9 Apr 2002 bre...@eu... wrote: > I think this is a healthy reading if you are interested in > multithreading implementations: > > http://supertech.lcs.mit.edu/cilk/manual-5.3.1.pdf > > Cilk homepage: http://supertech.lcs.mit.edu/cilk > > Cilk is a minimalistic extension of C that allows parallel > programming. Actual threads (or processes) are started by the > scheduler when a processor is idle, according to the > "work-stealing" principle. > > -- Denis. > > > _______________________________________________ > Nomen-dev mailing list > Nom...@li... > https://lists.sourceforge.net/lists/listinfo/nomen-dev > |