By: anon (anon.delete@this.anon.com), October 17, 2012 1:17 am
Room: Moderated Discussions
Paul A. Clayton (paaronclayton.delete@this.gmail.com) on October 16, 2012 11:09 am wrote:
> The discussion of lack of large-scale bandwidth in supercomputers brings to mind
> William Kahan's quote "The Fast drives out the Slow, even if the Fast is
> wrong."; perhaps something like "The Big drives out the Small, even if the Big
> is useless." might almost apply.
>
> Unfortunately for workloads that have
> demanding cross-node communication, "bandwidth is only money" is not much
> comfort when the cost of high all-to-all bandwidth is so great.
Exactly. This is why low bandwidth, high latency memory and communications is not the problem, but the *solution*. Together with caches and changed software assumptions, of course.
Declaring the solution to in fact be the problem and vowing to do away with it just leaves you holding the bigger problem. Handwaving about "streaming" or "optical" does not solve the problem. People need something that works now.
> The discussion of lack of large-scale bandwidth in supercomputers brings to mind
> William Kahan's quote "The Fast drives out the Slow, even if the Fast is
> wrong."; perhaps something like "The Big drives out the Small, even if the Big
> is useless." might almost apply.
>
> Unfortunately for workloads that have
> demanding cross-node communication, "bandwidth is only money" is not much
> comfort when the cost of high all-to-all bandwidth is so great.
Exactly. This is why low bandwidth, high latency memory and communications is not the problem, but the *solution*. Together with caches and changed software assumptions, of course.
Declaring the solution to in fact be the problem and vowing to do away with it just leaves you holding the bigger problem. Handwaving about "streaming" or "optical" does not solve the problem. People need something that works now.
