I’ve recently been focusing on CAE: who is using cutting-edge tools, why, where they see
improvements, where they still have issues. One theme that keeps coming up is that the
bottlenecks are no longer where they used to be. As compute power increases to the point
where enough money will accomplish almost anything in a reasonable timeframe, we’re now
looking at other parts of the computational cycle.
Back in June, the New York Times wrote about the new Roadrunner supercomputer being
built by IBM and the Los Alamos National Labs. For $133 million, you can buy the compute
power needed to perform 1.026 quadrillion calculations per second. To put this in
perspective, the NYT article has this tidbit: "Thomas P. D’Agostino, the administrator of the
National Nuclear Security Administration, said that if all six billion people on earth used
hand calculators and performed calculations 24 hours a day and seven days a week, it
would take them 46 years to do what the Roadrunner can in one day." Wow.
OK, so we can now calculate how nuclear weapons explode, how climate patterns affect us
and lots of other things. But what then? What happens to the output?
According to data supplied by Stan Posey, now at Panasas but at SGI when I first met him,
if it took 10 minutes to read the file containing the mesh and boundary conditions and and
35 minutes to write the results to disk for an analysis that 14 hours to calculate in the
1990s, the I/O took 5% of the total time. Now that’s a slow overall turnaround — one of these
per day was the limit and meant that analysis was usually done a specific points in a design
cycle. Today, that same analysis done on a high performance compute cluster might take
26 minutes — but the I/O still takes 45, or 63% of the total time. Much faster, so analysis is
likely more a "what-if", frequently used part of the design process — but there’s a lot of time
"wasted" that keeps the analyst from his results. [In full disclosure, Panasas’ products are
intended to streamline these I/O bottlenecks.]
But even if one could reduce these I/O bottlenecks just as CPU times have been
phenomenally reduced, we’re still left with post-processing. Output files these days run to
terabytes; doing any meaningful visualization on the results is still a massive process of
partitioning the output file, creating a human-intelligible output and sewing the parts together.
Will that be the next area for innovation in the CAE space?