By: forestlaughing (forestlaughing.delete@this.yahoo.com), October 16, 2012 9:13 am
Room: Moderated Discussions
Robert Myers (rbmyersusa.delete@this.gmail.com) on October 15, 2012 9:04 am wrote:
> > And you wouldn't know anything about self promotion, would
> > you
> Robert? ;)
> >
> If people recognize my name at all, it's because I have a few hobby
> horses that I have mostly stuck to. Given that I usually pick topics on which
> there is a tidal wave of mindless repetition opposed to what I am usually
> saying, it's perhaps a shame that I don't have even more talent for
> self-promotion, if that's what you want to call it. If you think I'm in the
> same league with IBM, LLNL, ORNL, and the University of Tennessee, I'm
> flattered. They have entire teams at each location. There is just one of
> me.
I don't mean to poke at you too much. Self promotion is one of the primary job functions of the academic.
Your argument is valid. Many applications are left behind in the rush to commodity processors, and huge-scale MPPs. However, I think you mischaracterize and underestimate the labs and vendors. No one is thrilled by the relative drop-off in memory and network bandwidth, compared to flops. Everyone would like machines that are lower latency, have gobs of bandwidth, are easier to program, and use very little power. It is also that case that the labs are aware that smaller machines, with higher per-node performance, help a number of applications, and many labs do buy such machines, they are just not the flagship system.
The NNSA labs have the luxury of buying supercomputers to run 8-10 applications. However, most of the rest of HPC buyers have a diverse user base, and would really like to offer revolutionary performance for all of their users. However they're just buying what they're able to buy, and meeting the needs of as many users as possible. To be sure, the staff at the big labs look at running linpack as an tiresome chore. They have to crow a little, about their accomplishments, and linpack is the only metric the news media is interested in reporting. They don't buy fast linpack machines in order to run linpack. They buy the best machine they can, for the largest number of their users as they can, given a finite budget, and then talk about linpack because that's the only industry-wide vocabulary.
> > And you wouldn't know anything about self promotion, would
> > you
> Robert? ;)
> >
> If people recognize my name at all, it's because I have a few hobby
> horses that I have mostly stuck to. Given that I usually pick topics on which
> there is a tidal wave of mindless repetition opposed to what I am usually
> saying, it's perhaps a shame that I don't have even more talent for
> self-promotion, if that's what you want to call it. If you think I'm in the
> same league with IBM, LLNL, ORNL, and the University of Tennessee, I'm
> flattered. They have entire teams at each location. There is just one of
> me.
I don't mean to poke at you too much. Self promotion is one of the primary job functions of the academic.
Your argument is valid. Many applications are left behind in the rush to commodity processors, and huge-scale MPPs. However, I think you mischaracterize and underestimate the labs and vendors. No one is thrilled by the relative drop-off in memory and network bandwidth, compared to flops. Everyone would like machines that are lower latency, have gobs of bandwidth, are easier to program, and use very little power. It is also that case that the labs are aware that smaller machines, with higher per-node performance, help a number of applications, and many labs do buy such machines, they are just not the flagship system.
The NNSA labs have the luxury of buying supercomputers to run 8-10 applications. However, most of the rest of HPC buyers have a diverse user base, and would really like to offer revolutionary performance for all of their users. However they're just buying what they're able to buy, and meeting the needs of as many users as possible. To be sure, the staff at the big labs look at running linpack as an tiresome chore. They have to crow a little, about their accomplishments, and linpack is the only metric the news media is interested in reporting. They don't buy fast linpack machines in order to run linpack. They buy the best machine they can, for the largest number of their users as they can, given a finite budget, and then talk about linpack because that's the only industry-wide vocabulary.
