Nonlinear systems and nonlocal supercomputing
|
Prev: Call for Papers Reminder (extended): The World Congress on Engineering WCE 2010
Next: Call to stop spamming here
From: MitchAlsup on 16 Mar 2010 15:27 On Mar 16, 12:05 pm, "Del Cecchi" <delcec...(a)gmail.com> wrote: > And I might add that "no magic, just a sufficiently large crossbar > switch..." is slight of hand at best. Consider a network of 10,000 > nodes. Is one to build a single crossbar (non blocking I presume) > that has 10,000 input ports and 10,000 output ports all running at > some large bandwidth? Note that such a crossbar would HAVE to be decompoase into a number of layers with fewer ports and more layers--such that sooner or later routing decisions are performed within a chip that has a 'reasonable' number of pins. > What is it supposed to do if two of them want > to talk to the same output? buffer things up? The only other option is to route conflicting request out some link that is not currently busy. The receiving 'bar will route it back (or forward) so that eventually it does get where it needs to go. The memory interconnect in the H.E.P actually did this. > Is there some sort of > router table or is this a single stage thing so routing is implicit in > the destination? The telephone network tells us that having more than one route from source to destination is better. > Does it do clock frequency matching? I suspect it simpy has to be frequency matched but phase insensitive. > As someone once said in a different context "the network is the > computer" In this case the network is sure to cost more than the computers within. Mitch
From: Del Cecchi on 16 Mar 2010 15:56 "MitchAlsup" <MitchAlsup(a)aol.com> wrote in message news:2bbd9e9d-8253-4e7c-bf9c-e409f2804f8f(a)19g2000yqu.googlegroups.com... On Mar 16, 12:05 pm, "Del Cecchi" <delcec...(a)gmail.com> wrote: > And I might add that "no magic, just a sufficiently large crossbar > switch..." is slight of hand at best. Consider a network of 10,000 > nodes. Is one to build a single crossbar (non blocking I presume) > that has 10,000 input ports and 10,000 output ports all running at > some large bandwidth? Note that such a crossbar would HAVE to be decompoase into a number of layers with fewer ports and more layers--such that sooner or later routing decisions are performed within a chip that has a 'reasonable' number of pins. > What is it supposed to do if two of them want > to talk to the same output? buffer things up? The only other option is to route conflicting request out some link that is not currently busy. The receiving 'bar will route it back (or forward) so that eventually it does get where it needs to go. The memory interconnect in the H.E.P actually did this. > Is there some sort of > router table or is this a single stage thing so routing is implicit > in > the destination? ----------------------------------- sorry for attributions being hosed. It used to just be R. Meyer's posts. I thought it was google's fault. maybe not. ------------------------------ That's what I thought. That is why I asked these sort of rhetorical questions in response to Larry's post where he said: "Finally, I want to point out that good global communications can be done, it is just expensive. No magic is necessary. All you need to do is build a sufficiently large crossbar switch. These can have modular components, and it is "just" an engineering problem. Of course it's costs go as N**2 in the number of nodes." A true any to any non blocking crossbar like Larry proposed is economically unfeasible and probably physically impossible to actually construct. Once you do multilayer routing it doesn't do what the critics want. >The telephone network tells us that having more than one route from >source to destination is better. Telephone company doesn't care much about latency >> Does it do clock frequency matching? >I suspect it simpy has to be frequency matched but phase insensitive. >> As someone once said in a different context "the network is the >> computer" In this case the network is sure to cost more than the computers within. Mitch
From: nmm1 on 16 Mar 2010 16:02 In article <80a5vfFaemU1(a)mid.individual.net>, Del Cecchi <delcecchi(a)gmail.com> wrote: >"MitchAlsup" <MitchAlsup(a)aol.com> wrote in message >news:2bbd9e9d-8253-4e7c-bf9c-e409f2804f8f(a)19g2000yqu.googlegroups.com... > >>The telephone network tells us that having more than one route from >>source to destination is better. > >Telephone company doesn't care much about latency Below 10 milliseconds, no. >>> As someone once said in a different context "the network is the >>> computer" > >In this case the network is sure to cost more than the computers >within. As in large SGI Origins. Good systems, but you paid for bleeding edge networking. I am sure you can think of other examples :-) Regards, Nick Maclaren.
From: Robert Myers on 16 Mar 2010 16:28 On Mar 16, 3:27 pm, MitchAlsup <MitchAl...(a)aol.com> wrote: > On Mar 16, 12:05 pm, "Del Cecchi" <delcec...(a)gmail.com> wrote: > > > And I might add that "no magic, just a sufficiently large crossbar > > switch..." is slight of hand at best. Consider a network of 10,000 > > nodes. Is one to build a single crossbar (non blocking I presume) > > that has 10,000 input ports and 10,000 output ports all running at > > some large bandwidth? > > Note that such a crossbar would HAVE to be decompoase into a number of > layers with fewer ports and more layers--such that sooner or later > routing decisions are performed within a chip that has a 'reasonable' > number of pins. > > > What is it supposed to do if two of them want > > to talk to the same output? buffer things up? > > The only other option is to route conflicting request out some link > that is not currently busy. The receiving 'bar will route it back (or > forward) so that eventually it does get where it needs to go. The > memory interconnect in the H.E.P actually did this. > > > Is there some sort of > > router table or is this a single stage thing so routing is implicit in > > the destination? > > The telephone network tells us that having more than one route from > source to destination is better. > > > Does it do clock frequency matching? > > I suspect it simpy has to be frequency matched but phase insensitive. > > > As someone once said in a different context "the network is the > > computer" > > In this case the network is sure to cost more than the computers > within. > That's easy to see, even without the appropriate level of technical expertise to back up such a declaration in detail. I'm not certain of the sequence of events that has led to the current ordering of priorities, but I'm fairly certain that, one by one, bureaucrats decided to hoist the white flag on this problem because the cost of facing it was prohibitive. Much easier to add processors, flops, and one-by-one routing chips that will undoubtedly result in very pretty pictures (one of which Del proudly displayed), but where the physics and mathematics are questionable, for reasons that I have attempted to explain. If you're going to build computers that require rewiring of the nation's electric grid and that siphon huge amounts of money from other possible and much more solid lines of research, then it makes no sense to be fudging the math unless you absolutely have no other choice. Nothing would please me more than to hear that it's all been looked at in detail and that what I think is so fundamental about nonlinear computation isn't so fundamental, after all. If such an argument has ever been advanced, I've never seen it. When I started into the turbulence business a long time ago, one suggested possibility is that boundary layer breakdown on airfoils in subsonic flight migrated forward from the sharp trailing edge, where vorticity shedding is what keeps airplanes aloft, and where all kinds of detailed microscopic physics are underway. I don't know what has happened to the suggestion, but there was no reason at the time to dismiss it out of hand. If the propagation of information forward were determined by artificial localization of differential operators, you would very unlikely *ever* unscramble the phsyics using a machine like Blue Gene or any of its brethren. Robert.
From: MitchAlsup on 17 Mar 2010 13:08
Robert, can you contact me via e-mail. I had a though I want to run down and cross check before posting again. Mitch |