The Argonne National Laboratory was keen to examine what advantages the cloud computing model could offer to the scientific research community. In a joint research effort with the National Energy Research Scientific Computing Center (NERSC), it created the Magellan cloud testbed project. The Argonne Leadership Computing Facility (ALCF) at Argonne National Laboratory is operated by the Leadership Computing Facility Division as part of the U.S. Department of Energy’s effort to provide leadership-class computing resources to the scientific community. Argonne was keen to examine what advantages the cloud computing model could offer to the scientific research community. In a joint research effort with the National Energy Research Scientific Computing Center (NERSC), it created the Magellan cloud testbed project. The aim was to explore a cloud solution that might allow existing projects to take advantage of a more powerful and flexible compute resources, and potentially reduce the pressure on its other high-performance computing (HPC) resources.

Pete Beckman, director of the Exascale Technology and Computing Institute at the Argonne National Laboratory, says, “The Magellan project is an experiment to find out what kind of science can be done in a cloud, and what kinds of changes we might have to make either to the technical platform or to the software. Our goal is to understand what kinds of scientific applications can run well in the cloud, and look at running them in the cloud on a grand scale.”

Robust performance
Magellan is based on the IBM System x iDataPlex dx360 M3 server, a, half-depth server optimized for maximum density and incredible efficiency. Argonne deployed 504 iDataPlex compute nodes, each with two quad-core Intel Xeon X5550 processors running at 2.66 GHz, providing a total of a 4,032 cores. Each node has 24 GB of DDR3 1333 MHz memory.

The iDataPlex solution also offers 200 expanded compute nodes, using the same processors but featuring 48 GB of DDR3 1066 MHz memory, a 200 GB Solid State Drive and a 1 TB local SATA disk per node. There are also 12 login and network service nodes in Magellan.

Argonne made use of the IBM Intelligent Cluster offering, including the physical installation of the cluster and system cabling by the IBM Cluster Enablement Team. All nodes, switches and other components were installed and integrated into the racks during manufacture, and the solution was shipped pre-racked for added speed of deployment. The IBM Cluster Enablement Team also handled software installation and system configuration on-site at Argonne.

Susan Coghlan, deputy division director at the Argonne Leadership Computing Facility says, “NERSC and Argonne have the IBM iDataPlex hardware for their core compute cloud testbed. It’s a very robust platform and it’s exactly what Argonne has on its production HPC cluster. We’re focused on the system software and the science that may be able to run on a cloud, as opposed to the hardware. We want the hardware to just stay up and run without us having to think about it.”

IBM delivered a complete turnkey solution, including high performance Mellanox ConnectX InfiniBand adapters to provide high bandwidth, low-latency interconnects between the iDataPlex compute nodes. IBM also set up the management nodes for the new solution and the entire surrounding network.

Pioneering solution
Magellan is testing how cloud technologies could make a direct contribution to the scientific community, by potentially enabling unprecedented flexibility, efficiency and speed of provisioning for high-performance computing resources.

The cloud model is potentially well suited to the study of genomics and meta-genomics in the field of biology. Unlike traditional supercomputing applications, which tend to use large parallel applications running on a parallel framework, the applications used by biologists are much more pipeline-driven.

Accelerating computation
The Intel Xeon 5550 processors used in the IBM iDataPlex employ Intel’s Turbo Boost technology, which enables some cores to be temporarily overclocked to overcome bottlenecks in processing. When demand falls, the cores can be underclocked to reduce their energy consumption and heat output.

“From a user perspective,” says Beckman, “what he wants is something that runs blindingly fast and he never knows that it’s there. The IBM architecture allows data to be transferred back and forth from the CPU to memory, so computation can continue at the same time that the system is moving data to and from the nodes.”