It seems that everything else is heading to the cloud these days, so it shouldn't be a surprise to see even the high-performance computing (HPC) community getting in on the act.
But the question remains: Is the cloud robust enough to handle the massive compute cycles and high throughput needed for an effective HPC environment?
SGI certainly thinks so and has unveiled its new Cyclone platform to prove it. Built on the Intel-powered Altix system, Cyclone is the first major announcement since Rackable Systems bought out the old SGI last spring and adopted the company name. The service will be available on a range of Xeon- or Itanium-based Altix components, such as scale-up cluster packages or the XE hybrid system, and complemented with NVidia or AMD GPUs and Tilera accelerators for heavy workloads. Users will be able to access the system either as a SaaS service, with applications ranging from the openFOAM fluid dynamics package to the BLAST and FASTA systems geared toward life sciences workloads, or as an intrastructure-as-a-service platform to house their own apps. Users also have access to the InfiniteStorage cloud for near-line or archival storage. Prices start at a relatively modest 95 cents/HPC core/hour and 20 cents/GB/month for storage.
While SGI has built itself around HPC platforms over the years, it is not the only one targeting new cloud services. A company called Penguin Computing already has six months' worth of experience running the Penguin on Demand (POD) service and is said to have gathered more than 2,000 users so far. The company forgoes the use of virtualization, instead offering dedicated hardware that provides faster throughput than shared systems and can be more easily optimized for the parallel, multi-threaded applications preferred by HPC users.
Demand for HPC on the cloud, particularly from large institutional users, is drawing specialty organizations together to fill the need. Management firm Platform Computing, for example, recently linked up with HPC firm Instrumental Inc. to provide a cloud-based HPC service aimed at U.S. government agencies. The service is based on Platforms' ISF and ISF Adaptive Cluster systems and is designed to supplement legacy heterogeneous HPC operations, as well as related hardware and virtual environments, giving agencies the ability to scale up operations without substantial upfront capital costs.
The ability to expand capabilities without investing in new hardware has always been a chief selling point for the cloud. But now it seems that price pressures may bring the cost of HPC cloud services down as well. A company called Univa announced this week that it was making its UniCloud and Grid MP services available on Amazon EC2's Spot Instances platform, which gives users the chance to bid for capacity auction-style. Through the UniCloud policy engine, users can set their price level and tap into EC2 when resources are available at low cost. If bidding is driving the cost up, the system can shift workloads to other resources.
At the moment, it appears that most HPC organizations are using the cloud to supplement their existing services. That's to be expected since these applications typically deal with immense amounts of data and phenomenal levels of computing power.
But if the typical enterprise is still unsure of the security and reliability of the cloud, it will only be that much longer before the HPC industry is ready to commit to a fully clouded solution -- if ever.