[Beowulf] K Computer built for speed, not use

[Lux, Jim (337C)]

Actually, I think it's summed up here:
"Many, if not most, users are just happy to have a faster computer"
" Forming a consensus among the high-level users would have taken time, "leaving end users like me waiting for many years," he says"

Yep.. perfect is the enemy of done.

And I note that the potential griper Makino-san is still lining up to use it, non-optimum architecture and all.

Reading the slides.. while NWT and CP-PACs he rates as good or very good, they had a single science goal.. so yeah, of course it worked well for that one user: they could optimize.
Any general purpose system will inevitably underperform for some people, and many might argue that the art of managing such a project is making sure everyone squawks equally loud about how the stake is being driven into their heart.

One also wonders if Makino-san is angling to be named the "single person as leader" for exascale...

(I am also cautioned about the possibility of poor translation.  What might be an idiomatic phrase in Japanese might mean something different when translated into English.)

Jim Lux


-----Original Message-----
From: beowulf-bounces at beowulf.org [mailto:beowulf-bounces at beowulf.org] On Behalf Of Mark Hahn
Sent: Wednesday, October 10, 2012 4:31 PM
Cc: Beowulf Mailing List
Subject: Re: [Beowulf] K Computer built for speed, not use

> http://www.sciencemag.org/content/338/6103/26.full?rss=1
...
> Interesting claim. What kind of architecture structure would benefit 
> Linpack and would hinder real-world applications?

my guess is they don't like vector.

the quote appears to be from Jun Makino, the GRAPE guy (so qualifies as "sour grapes!).  there is a hint of critique in http://jun.artcompsci.org/talks/oookayama20120116.pdf
(which if I read between the lines is saying that for his kind of astrophysics, he wants accelerator-type architectures, which differ significantly from vector archs in their relation of cpu and memory.  the "1+3 architectures" table seems to show a desire for dramatically lower B/F (bytes of memory BW per flop?) as well as much lower M/F.

as for linpack being a bad benchmark, that's just bullshit.
it's a benchmark.  it's not your application.  it does a good job of telling us about a form of performance that is well-understood.
yes, you can make a very good guess at HPL performance if you know ncpus, peak FP rate and the interconnect performance - but conversely, a benchmark which is unpredictable is nothing to brag about!

as for the criticism of K's process, well, making sausage is ugly.
everything about a big project is sausage-like, and K is a fairly remarkable success given the range of issues it had to span.
even ignoring the politics and finance, a sparc chip (!) that does very wide SIMD, with a memory system to support it, a cooling design to keep it going, and interconnect to scale.

the only thing that pains me about the whole thing is that I don't guess all the lessons learned will be propagated or leveraged.  mini-K will not be coming to a center near you.  there won't be an commodity chip that gets "now with added K sauce"...
_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf