Every now and then, you read something that just pushes all your buttons. At breakfast this morning, I was reading an article in Nov 17 Wired magazine (because Marla cut the daily newspaper delivery which used to be my morning staple, but I digress). The article is “Into the Vortex”, sorry I couldn’t find an online link. It talks about the May 2013 tornado that killed storm chasers Tim Samaris, his son Paul, and Carl Young. Tornado nut that I am, I remember it vividly like it was yesterday.
The article went on to talk about how a researcher had been trying to simulate tornado formation on computers for years and finally got a chance to work with the supercomputer Blue Waters at the University of Illinois. It takes that kind of power to do the computations necessary. Using soundings from a similar EF-5 tornado from 2011, he was finally able to simulate a tornado.
Since we’re doing a lot of AWS (Amazon Web Services) work these days and working with Big Data and machine learning, Blue Water’s power piqued my interest. Just how big is it, and could it be re-created on AWS? Looking at computational power alone, Blue Waters can process 13 quadrillion operations a second, or 13 petaflops.
That got me searching for an AWS equivalent story, and I came across this Ars Technica article, 18 hours, $33K, and 156,314 cores: Amazon cloud HPC hits a “petaflop”. which claimed a theoretical peak speed of 1.21 petaflops. ~10 times slower, but more importantly, really geared for a different kind of workload. The AWS Compute Cluster is great for parallel processing, millions of tasks that can be performed independently because the processing nodes are geographically distributed and don’t need fast intercommunications. For the tornado research, only something like Blue Waters would be up to the task as the nodes “had to communicate the movement of air between its 27,000 nodes, processing at .2 second time steps. With all that power, it still needed almost a full day to perform the simulation.
So there’s a peek into my morning coffee routine 🙂