Like many organizations that rely on industrial-strength datacenters, the US Department of Energy (DOE) would like to know if cloud computing can make its life easier. To answer that question, the DOE is launching a $32 million program to study how scientific codes can make use of cloud technology. Called Magellan, the program will be funded by the American Recovery and Reinvestment Act (ARRA), with the money to be split equally between the the two DOE centers that will be conducting the work: the Argonne Leadership Computing Facility (ALCF) at Argonne National Laboratory and the National Energy Research Scientific Computing Center (NERSC) at Lawrence Berkeley National Laboratory.

One of the major questions the study hopes to answer is how well the DOE's mid-range scientific workloads match up with various cloud architectures and how those architectures could be optimized for HPC applications. Today most public clouds lack the network performance, as well as CPU and memory capacities to handle many HPC codes. The software environment in public clouds also can be at odds with HPC, since little effort has been made to optimize computational performance at the application level. Purpose-built HPC clouds may be the answer, and much of the Magellan effort will be focused on developing these private "science clouds."

The bigger question, though, is to find out if the cloud model in general is applicable to high performance computing applications used at DOE labs and can offer a cost-effective and flexible approach for researchers. According to ALCF director Pete Beckman, that means getting the best science for the dollar. In a cloud architecture, the virtualization of resources usually translates into better utilization of hardware. In the HPC realm though, virtualization can be a performance killer and utilization is often not the big problem it is in commercial datacenters where hardware is typically undersubscribed. Perhaps of greater interest for HPC users is the ability to fast-track application deployment by taking advantage of the cloud's ability to encapsulate complete software environments.

"There are a lot users who spend time developing there own software inside their own software stack," says Beckman. "Getting those running on traditional supercomputers can be quite challenging. In the cloud model, sometimes these people find it easier to bring their software stack with them. That can broaden the community."

The entire range of DOE scientific codes will be looked at, including energy research, climate modeling, bioinformatics, physics codes, applied math, and computer science research. But the focus will be on those codes that are typically run on HPC capacity clusters, which represent much of the computing infrastructure at DOE labs today. In general, codes that require capability supercomputers such as the Cray XT and the IBM Blue Gene are not considered candidates for cloud environments. This is mainly because large-scale supercomputing apps tend to be tightly coupled, relying on high speed inter-node communication and a non-virtualized software stack for maximum performance.

Most of the program's $32 million will, in fact, be spent on new cluster systems, which will form the testbed for Magellan. According to NERSC director Kathy Yelick, the cluster hardware will be fairly generic HPC systems, based on Intel Nehalem CPUs and InfiniBand technology. Total compute performance across both sites will be on the order of 100 teraflops. Yelick says there will also be a storage cloud, with a little over a petabyte of capacity. In addition, flash memory technology will be used to optimize performance for data-intensive applications. The NERSC and ALCF clusters will be linked via ESnet, the DOE's cutting-edge 100 Gbps network. ESnet was also a recipient of ARRA funding, and will be used to facilitate super-speed data transfers between the two sites.

One of the challenges in building a private cloud today is the lack of software standards. However, the Magellan work will employ some of the more popular frameworks that have emerged from the cloud community. Argonne, for instance, will experiment with the Eucalyptus toolkit, an open-source package that is compatible with Amazon Web Services API. The idea is to be able to build a private cloud with the same interface as Amazon EC2.

Apache's Hadoop and Google's MapReduce, two related software frameworks that deal with large distributed datasets, will also be evaluated. Like Eucalyptus, Hadoop and MapReduce grew up outside of the HPC world, so currently there's not much support for them at traditional supercomputing centers. But the notion of large distributed data sets is a feature of many data-intensive scientific codes and is a natural fit for cloud-style computing.

The other aspect of the Magellan effort has to do with experimentation of commercial cloud offerings, such as those from Amazon, Google, and Microsoft. Public clouds, in particular, are attracting a lot of interest due to their ability to offer virtually infinite capacity and elasticity. (Private clouds, because of their smaller size, tend to be seen as fixed resources.) Just as important to the DOE, a public cloud has the allure of offloading the development and maintainence of local infrastructure to someone else.

"Will it be more cost effective for a commercial entity to run a cloud, and presumably make a profit on it, than for the DOE to run their own cloud?" asks Yelick. "That is going to be one of the questions most challenging to answer."

Some DOE researchers are already giving public clouds a whirl. Argonne's Jared Wilkening recently tested the feasibility of employing Amazon EC2 to run a metagenomics application (PDF). The BLAST-based code is a nice fit for cloud computing because there is little internal synchronization, therefore it doesn't rely on high performance interconnects. Nevertheless, the study's conclusion was that Amazon is significantly more expensive than locally-owned clusters, due mainly to EC2's inferior CPU hardware and the premium cost associated with on-demand access. Of course, given increased demand for compute-intensive workloads, that could change. Wilkening's paper was published in Cluster 2009, and slides (PDF) are available on the conference Web site.

The Magellan program is slated to run for two years, with the initial clusters expected to be installed sometime in the next few months. At NERSC, Yelick says the hardware could arrive as early as November, and become operational in December or January. Meanwhile at Argonne, Beckman is already running into researchers who can't wait to host their codes on the Magellan cloud. "They're lined up," he says. "They keep coming down to my office asking when it will be here and how soon they can log in."