[ExI] simulation

[J. Andrew Rogers]

On Dec 15, 2007, at 10:02 AM, Bryan Bishop wrote:

> On Saturday 15 December 2007, J. Andrew Rogers wrote:
>> Are you using magic fiber?  Is the speed of light different in your
>> glass than in the rest of the universe?
>
> If all of the nodes are local, communication at c will work well  
> enough.
> Wasn't that Eugene's point?


Even very locally, the latency difference between fiber Ethernet and  
specialized interconnects can still be measured in orders of magnitude  
such that computational problems that are practical on the latter are  
not cost effective on the former.

Your enemy here is latency and many problems (including the one being  
discussed here) have parallelizability that is a very sensitive  
function of latency.  It does not matter how many machines you have  
locally if the interconnect latency is so high (where "high" may be  
tens of microseconds) that you can only use a small fraction of them  
on your problem.  This is the reason many supercomputing clusters will  
pay more for their cluster interconnect than they will on the actual  
compute nodes; having a lot of CPUs does not help much if your network  
latency is too high to parallelize the problem over that many nodes.


J. Andrew Rogers