ScaleMP, a Cupertino, CA based company, takes multiple physical machines and turns them into a single virtual image for high performance computing. Unlike Beowulf, which runs a single application across multiple machines, ScaleMP creates a single image with one operating system and the combines resources of all the servers.
Why would you want to use ScaleMP instead of buying a single supercomputer?
According to ScaleMP CEO Shai Fultheim, you can save money by combining multiple machines instead of buying a single large one. The cost of servers is non-linear: a two socket server may cost $5,000, a four socket server $25,000 and an eight socket server $80,000.
The average ScaleMP customer is running 14 servers with 28 sockets. Some have virtual machines with hundreds of sockets. Applications include algorithmic trading, bioinformatics, complex simulations and genome research.
Fultheim says one of ScaleMP’s clients needs to run an analytics job once a quarter. This job requires hundreds of gigabytes of memory. If this company didn’t have ScaleMP it would need to have one big machine that it only used once per quarter. Instead, it can pool the resources of several of its servers to run the job, and then use those machines for other tasks the rest of the quarter.
ScaleMP is also experimenting with the possibility of running virtual machines, such as VMware or Xen servers, on top of one of its own virtual machines. The company has completed a technology preview, but Fultheim it’s not ready for production yet. Other future plans include taking advantage of GPUs.
ScaleMP was founded in 2003. Its first deployment was in 2006, and it had three customers by the end of that year. By the end of 2010, it had 220 customers. According to Inside HPC the company secured series E funding in 2010 from Evolution Venture Capital and “a few private investors.”