SC10 Trip Report

[#1]

This was the 23rd annual conference on high performance computing (HPC), networking, storage, and analysis. Josh Simons (from the VMware CTO's office) and I have been working for the past few months on evangelizing the advantages virtualization brings to HPC applications, as well as evaluating various technical issues. My main motivation in attending was to learn more about current HPC apps and the level of computational resources they require. I was also very curious to see how aware attendees were of virtualization, or if it was on their radar screens at all.

The conference took place at the Convention Center in New Orleans. This is an excellent venue for large conferences. More than 10 thousand attendees were uncrowded in half the building, with another large conference in the other half. Keynotes took place in the amphitheatre-like main auditorium, unlike the usual arrangement in a flat-floored hall. The show floor was immense, with over 300 booths displaying and demoing state-of-the-art hardware and software. One of the "Disruptive Technologies" booths was run by Josh and four VMware partners (UnivaUD, Oak Ridge National Laboratories, Ohio State, and Deopli). They had a fair number of people come by to talk about virtualization and HPC. But after several VMworlds, it was a bit of shock to be associated with one of the "little guys" on the show floor. Among the big and popular booths was Nvidia, who showed off the use of their technology as GPGPU (general purpose GPU). With 480 cores on their latest chip, "Fermi" has the potential to deliver dozens of times the floating-point performance as conventional CPUs while consuming as little as 1/10 of the power. Intel attracted a lot of attention by getting Drew Brees, home-town hero, to agree to be interviewed at their booth.

I spent most of my time attending talks. Clayton Christensen in the main keynote hardly mentioned HPC (forget about virtualization), but had some fascinating insights about the rise and fall of successful companies. In particular, the same management team can be (and often is) responsible for both. As on the show floor, GPGPU was often the center of attention. But there were detractors who noted that because of memory bandwidth and latency issues, many applications do not run well on GPGPU machines. Others noted that one of the main impediments to the efficient use of highly parallel machines is writing and debugging programs for them, and that dealing with GPGPU interfaces makes this even harder. These are all important issues for VMware as well. Low-latency interconnects was another recurring theme. IBM showed a 192-node Power7-based cluster that is designed for less than 1 microsecond latency across the cluster and 8 TB/s networking bandwidth out of each node. D.E. Shaw Research described Anton (no relation to German lovers), a machine with sub-microsecond latency across 12 hops of the cluster network. This is a special-purpose machine with fixed communication patterns and other simplifications that are suitable for molecular dynamics simulations. I saw only one talk that had significant virtualization content. This was about increasing TCP throughput in Xen VMs. I may be biased, but it seemed to just involve avoiding scheduling latency by immediately acknowledging packets in the hypervisor instead of waiting for the guest. The authors did acknowledge that this is not always safe.

I especially enjoyed the application-oriented talks. These ranged from climate simulation (several of those) to FFTs. Many speakers seemed to compete to see who had to store the most data. This was topped by a description of the next generation radio telescope array, where the storage of the raw data is considered to be out of the question. Instead machines running at 300 teraflops are required to reduce incoming data in real time so that only mere petabytes need to be stored.

No trip to New Orleans would be complete without a visit to the French Quarter. After dinner at a very nice West African restaurant, a few of us walked down Bourbon Street. So many bands were playing in the various establishments that you could always hear at least four at any time. While the street itself was packed with people, the bars and strip clubs on either side were not so crowded. Apparently after experiencing the thrill of supercomputers, nothing else is very exciting.