Amazon.com, the retailer who wants to offer everything, is now adding supercomputers to its inventory. More precisely, its Web Services LLC announced on Tuesday Cluster Compute Instances for Amazon EC2, specifically designed for on-demand use for high-performance computing.
The company said the pay-as-you-go service is designed for customers who have "complex computation workloads such as tightly coupled parallel processes," or who have applications that are particularly dependent on network performance.
Up to 10 Times Throughput
Businesses needing access to this much computing power can do so with the ability to scale on demand, make no upfront investments, and pay only for the service and time used. Amazon said Cluster Compute Instances offers the same kind of functionality as other EC2 services, but substantially more CPU power.
Depending on the use, the service could provide up to 10 times the network throughput of the largest current Amazon EC2 instance types. Cluster Compute Instances can themselves be grouped into clusters, to get low-latency network performance that node-to-node communication might require.
The company said its new service eliminates the cost and complexity of acquiring and operating in-house computer clusters, enables innovation by offering access to such resources within minutes, allows resources to be scaled to the workload, and allows businesses more flexibility in resources.
So that a developer can tune an application by compiling for a specific architecture, Cluster Compute Instances will indicate the processor architecture to be used.
Peter De Santis, Amazon EC2 general manager, told news media that businesses and researchers have used EC2 to run "highly parallel workloads" such as genomics sequence analysis and automotive or financial modeling, but have been telling Amazon that they needed better network performance. That led to the new offering.
As Low as 56 Cents
He said that, in pre-production tests, an 880-server subcluster achieved 41.82 TFlops on a LINPACK test run. As part of the tests, the National Energy Research Scientific Computing Center at Lawrence Berkeley National Laboratory test-drove its high-performance computing applications on Cluster Compute Instances. Lawrence is the primary high-performance computing facility for scientific research sponsored by the U.S. Department of Energy.
Keith Jackson, a Lawrence computer scientist, said benchmark tests indicated the laboratory's applications ran 8.5 times faster on Cluster Compute Instances than in previous EC2 instance types.
Amazon also said that MathWorks, a developer and supplier of software for technical computer and model-based design, is allowing its customers to use its MATLAB and Parallel Computing Toolbox on desktops, and to scale data-intensive problems up to Cluster Compute Instances, which runs MATLAB Distributed Computing Server.
Amazon has said that each node will have a pair of Intel Xeon X5570s, running at 2.93 GHz and with 8M of on-chip cache memory, connected by 10-gigabyte-per-second Ethernet interfaces. The cost is reported to be about $1.60 per hour per instance. If purchased in bulk for up to three years, the cost could be as low as 56 cents per hour per instance.