via Sun’s HPC Watercooler
Does mankind benefit? Maybe.
Primeur Weekly is running an article on the report of the HPC in Europe Taskforce (HET), an initiative of eleven European countries. There looks to be some sound reasoning here, for example the desire to foster a robust European HPC ecosystem
The HPC ecosystem can be described in terms of a performance pyramid. At the top is a small number of top-of-the line supercomputer systems, funded through national source, with additional European funding. The middle layer of the pyramid consists of a number of national and regional supercomputers. These still should be powerful supercomputers being able to run all the load below petaflop/s level…The bottom of the pyramid consists of local supercomputers that should enable the development of a strong competence base of computational scientists.
Fair enough. There’s also a focus on software, which is what all the kids on this side of the pond are talking about these days. But then there’s this:
European computational science needs computing capacities that match the fastest systems in the USA and Japan, HET says. It is an intolerable situation for the competitiveness of European computational science that this currently is not the case.
At least the press around this looks like another HPC arms race. In fact, the HET’s number 1 recommendation is to procure “extreme computing power.”
1. Recommendation for the development and operation of a “top end” infrastructure. HET recommends establishment of a small number of European HPC facilities to provide extreme computing power – exceeding petaflop capability – for the most demanding computational tasks.
Enough with the FLOPS already: you guys are going to spend, I don’t know, hundreds of millions of dollars, for machines that will give you at most 5% of peak on a diverse workload.
Why not set an example for the rest of the HPC community and put a focus on usability, productivity, and getting work done at the very top of your pyramid?
The bottom line is that the US and Japan are “solve it with money” cultures, so Europe probably won’t ever win in a FLOPS arms race. But why try? Lead on getting actual work done and beat the pants off the HPC superpowers with a resurgence in your own national competitiveness.
Just remember to keep an eye out in back as we come up fast from behind. We usually only have to pick up a red horseshoe once to know it’s hot.
It’ getting hard for self-respecting big nations/regions to resist the prestige of the “Petaflop Club,” despite the expensive member fees. Horst Simon, NERSC/LBL, is eloquent on the dangers of focusing on the first peak/Linpack petaflop, on grounds that this could convince funders that the battle is over, when the real work of getting to a usable petaflop would just have started and might take another 8-10 years. 15 years ago, Cray Research used to host annual “(sustained) gigaflop performance awards” that were based on real-world apps performance. The HPCChallenge benchmarks get part of the way to reviving real-world metrics, though unfortunately the annual awards are based on the optimized results that allow/encourage gaming the system. The real-world metrics need to keep up with the scalable architectures; otherwise, we risk escalating the trend toward increasingly inefficient, “flopsided” supercomputers.
Hey Steve! Thanks for posting…I wonder if we can get the sustained performance awards started again? We could even host them here at insideHPC.com to keep them “vendor neutral”.
I think another big step center directors could take is to stop putting their machines on the Top500 at all, focusing instead on something else (HPCChallenge?). It would be symbolic at first, but if enough of us did it it could get a conversation going about measures that are more meaningful.
IMO the problem is not with the Top500/Linpack itself, which is a very useful lowest-common-denominator metric for comparing many types of computers. The problem is that this metric does not come with an explicit “warning label” that clearly describes its intended purpose and the limits of its usefulness. Absent this warning, it is widely and wrongly perceived outside of the HPC industry as the ultimate measure of HPC system performance. The HPCchallenge tests are a major step forward. Unfortunately, a single test with a single number (Top500/Linpack) is vastly more compelling for marketing purposes.