Not all supercomputing rides on InfiniBand or proprietary interconnects. For technical applications that decompose neatly into loosely-coupled threads, a big cluster with vanilla Gigabit Ethernet does just fine. The persistence of Ethernet on the TOP500 attests to the interconnect’s continued viability on big clusters. On the latest June list, GigE is being used on 284 of the top systems, which is actually slightly up from the 273 recorded in November 2007. But as clusters scale out into hundreds or even thousands of nodes, Ethernet infrastructure can grow into a complex burden of cables and multi-layer switches.
The top Ethernet system on the TOP500 list — at number 58 — is the new ATLAS cluster at the Max Planck Institute for Gravitational Physics in Germany. Installed earlier this year, the ATLAS system is being used in the Institute’s quest to detect gravitational waves — one of the big prizes remaining in physics. A gravitational wave is a fluctuation in the curvature of space-time that is theorized to occur as the result of cosmic events in the early universe, or more recently, from the extreme gravitational fields generated by neutron stars and black holes. First predicted by Albert Einstein in 1917 as part of his General Theory of Relativity, gravitational waves have never been directly measured. Through the use of large arrays of laser interferometers deployed in the U.S., Italy and Germany, it is hoped that evidence of the elusive wave will be discovered.
Because the effect of gravitational waves are so subtle here on Earth, very large quantities of data must be collected, and enormous computational power must be brought to bear to prove their existence. It is hoped that the ATLAS system will provide a platform to help move this effort forward. The 32.8 teraflop (Linpack) machine is made up of 1,342 single-socket compute nodes, occupying 32 racks.
Each ATLAS compute server has a 2.4 GHz Intel quad-core Xeon processor and communicates with the rest of the system via a 1 Gigabit link to a top-of-rack Woven TRX 100 Ethernet switch, which acts as a GigE aggregator with four 10 GigE uplinks. The uplinks funnel the server data to the 144-port 10 GigE Woven EFX 1000 core switch. Since the configuration is not over-subscribed, non-blocking Ethernet communication is provided for each server.
Because of the amount of data involved in gravitational wave analysis, the ATLAS compute servers are hooked up to 1.3 petabytes of external storage. The storage consists of 42 separate file nodes, 30 of which are GigE-linked servers connected via another TRX 100; the other 12 are 10 GigE-connected Sun Microsystems “Thumper” file servers directly hooked into the EFX 1000 core switch. An additional 500 GB of direct-connected storage is provided on each compute node. The CPU on any server can access the local disk storage on any other server as well as the central storage nodes.
Unlike more tightly-coupled MPI codes, analysis of gravitational wave data is an embarrassingly parallel application that lends itself to a server farm type set up. Each node is involved in very data-intensive computations, but node-to-node communication is minimal. Most of the data communication takes place between the compute nodes and the storage.
Because of the highly parallel nature of the code and the reliance on low latency I/O communications, the more granular, single-socket servers were the best fit for the application. Bruce Allen, director of the Max Planck Institute for Gravitational Physics in Hannover, Germany, who led the specification ATLAS system, determined that even at the computational scale of the ATLAS system, a Gigabit Ethernet interconnect was the logical choice. “Something like InfiniBand or Myrinet would have been overkill for this kind of application,” he said.
What he really liked about Woven’s solution was how well designed and how cost-effective it was, and also how easily it scaled up to the 1,000-plus-node cluster he had in mind. Since the EFX 1000 incorporates 144 10 GigE ports, this single core switch, along with the TRX edge switches, supported compute and storage communication for the entire cluster. Another attraction of the Woven technology is its ability to dynamically determine the optimal path for the data. The vSCALE chip in the switch is constantly monitoring latency of the active and alternative paths in the Ethernet fabric. If it finds an alternative path with lower latency, the hardware redirects traffic to take advantage of the faster route. This is especially advantageous when all the nodes are accessing both central storage and local disks on the other nodes. According to Allen, the Woven hardware was better designed and more flexible than any other Ethernet solution they looked at.
“What is remarkable about the ATLAS cluster is that we were able to take the lead very cost-effectively with a creative combination of more processors at lower clock rates and a higher Ethernet switching efficiency,” explained Allen in a press release on Tuesday. “Woven’s 10 Gigabit Ethernet Fabric switch is able to deliver sustained performance at an impressive 64 percent of the theoretical peak. The HPC Linpack experts we consulted tell us that they have never seen such a high level of Ethernet efficiency on such a large cluster. Without the Woven switch, ATLAS would not be the world’s fastest Ethernet cluster. It’s that simple.”
Allen has also helped develop an even larger system that is being used to process the gravitational wave data. This one is also Ethernet-based, but communicates at sub-GigE speeds. The [email protected] project is a distributed grid of personal computers, and like its ATLAS sibling, is used to crunch some of the same laser interferometer data collected from around the world.
According the Allen, the current Einstein grid represents over 150 teraflops of computing power and adds about 2,000 new personal computers each day. Like the larger [email protected] and [email protected] projects, [email protected] relies on the kindness of strangers to donate spare PC cycles for the advancement of science. And while not as efficient as the Institute’s ATLAS supercomputer, the grid offers a lot of extra capacity for wave calculations. Between ATLAS and [email protected], another mystery of the universe may finally be revealed.