Penguin Puts High-performance Computing in the Cloud

Penguin Computing, which builds high-performance Linux clusters for tasks like weather modelling and product design, is taking its business into the cloud.

The company on Tuesday launched Penguin on Demand, or POD, a pay-as-you-go service that provides access to a HPC cluster built by Penguin and hosted at a data center in Utah.

The service is aimed at startups such as biotech companies that need HPC capacity but can’t afford their own supercomputer, and larger companies whose scientists and engineers need more capacity than their data centers can provide.

The POD service will compete with Amazon Web Services (AWS), which has signed up some biotech customers for its EC2 Elastic Compute Cloud, and with on-demand offerings from IBM and several smaller companies.

Penguin CEO Charles Wuischpard argued that AWS isn’t ideal for HPC work because of the performance overhead from using virtualization, and because customers don’t know where the servers they have been allocated are located. If they are far apart, that can cause latency that impairs performance for some HPC jobs.

“In high-performance computing you’re sending messages back and forth many times per second, so that latency adds up,” said Joshua Bernstein, a Penguin software engineer.

Penguin isn’t offering performance guarantees, however, except to say that its optimized Linux clusters will run HPC code faster than EC2. Wuischpard said the company can offer service-level guarantees to customers who require them.

It also hasn’t published a price list, saying only that it will be “at or below” Amazon’s rates. Amazon’s pricing starts at US0ドル.10 per virtual machine image per hour. Penguin will charge per CPU core per hour, and also on a monthly subscription basis.

William Fellows, a U.K.-based analyst with The 451 Group, said many traditional HPC customers have been experimenting with Amazon and other cloud-based services. “The Penguins of the world need to have an answer for that,” he said.

Penguin has a good track record designing specialized HPC clusters, he noted. Its customers include The National Weather Service, Caterpillar and Sandia National Laboratories, as well as Web 2.0 companies such as Digg and Plaxo.

Fellows agreed there can be a performance overhead with Amazon’s virtualized, highly distributed architecture. “But at the same time, you can go and rent an additional 2,000 CPUs for not very much money.”

The POD service is delivered from a cluster of Penguin’s Relion Servers with Xeon 5400 processors, with 4GB of memory per core, and an Infiniband or Gigabit Ethernet network fabric. The system is managed by Penguin’s Scyld cluster software and customers access it via a SSH (Secure Shell) interface.

The system in Utah has just over 1,000 CPU cores, relatively small to begin with, but Penguin will add capacity over time. It says it has two customers up and running today, both in bionformatics, and some pilot tests underway.

Customers who need to load large amounts of data into the system can ship it on disks, as they do with Amazon Web Services. Other concerns may include security and software licensing, since many commercial vendors charge by the core to use their products.

Penguin is one of several HPC companies branching out to find new business. Last week, Cluster Resources changed its name to Adaptive Computing and expanded its clustering software to support virtualized enterprise applications.

Penguin wants eventually to bring “HPC to the masses,” Wuischpard said. The POD service might appeal to customers who haven’t considered HPC before, he said, like small hedge-fund shops or even art students who want to render video.