Amazon EC2 has a new Linux-based instance type designed for HPC applications and other demanding network-bound applications. It’s called Cluster Compute Instances.

Treading on Penguin and SGI’s toes, both of which have HPC clouds, it claims complex computational workloads such as tightly coupled parallel processes and applications sensitive to network performance will see the same high compute and networking performance delivered by custom-built infrastructures but with the cloud’s elasticity.

Amazon says the new widgetry provides more CPU than any other EC2 instance. Customers can also group Cluster Compute Instances into clusters for the low-latency network performance required for the tightly coupled, node-to-node communication typical of many HPC applications.

Cluster Compute Instances are also supposed to provide significantly increased network throughput for applications that do network-intensive operations. Amazon figures that applications can see up to 10 times the network throughput of its largest EC2 instance types depending on usage patterns.

It said that in one of its pre-production tests, an 880-server sub-cluster achieved 41.82 TFlops on a LINPACK test run, which is outstanding.

Lawrence Berkeley National Lab, a big HPC facility, said the benchmarks it ran found that its HPC applications ran 8.5 times faster on Cluster Compute Instances than other EC2 instance types.

Prices start at $1.60 an hour for two quad-core machines. Reserved pricing runs 56 cents an hour with a one-year fee of $4,290 or a three-year fee of $6,590.

See http://aws.amazon.com.