This is one in a series of posts on when and where to use a distributed architecture design in your organization's computing needs. You can find the main post here: http://blogs.msdn.com/b/buckwoody/archive/2011/01/18/windows-azure-and-sql-azure-use-cases.aspx
High-Performance Computing (also called Technical Computing) at its most simplistic is a layout of computer workloads where a “head node” accepts work requests, and parses them out to “worker nodes'”. This is useful in cases such as scientific simulations, drug research, MatLab work and where other large compute loads are required. It’s not the immediate-result type computing many are used to; instead, a “job” or group of work requests is sent to a cluster of computers and the worker nodes work on individual parts of the calculations and return the work to the scheduler or head node for the requestor in a batch-request fashion.
This is typical to the way that many mainframe computing use-cases work. You can use commodity-based computers to create an HPC Cluster, such as the Linux application called Beowulf, and Microsoft has a server product for HPC using standard computers, called the Windows Compute Cluster that you can read more about here.
The issue with HPC (from any vendor) that some organization have is the amount of compute nodes they need. Having too many results in excess infrastructure, including computers, buildings, storage, heat and so on. Having too few means that the work is slower, and takes longer to return a result to the calling application. Unless there is a consistent level of work requested, predicting the number of nodes is problematic.
Recently, Microsoft announced an internal partnership between the HPC group (Now called the Technical Computing Group) and Windows Azure. You now have two options for implementing an HPC environment using Windows. You can extend the current infrastructure you have for HPC by adding in Compute Nodes in Windows Azure, using a “Broker Node”. You can then purchase time for adding machines, and then stop paying for them when the work is completed. This is a common pattern in groups that have a constant need for HPC, but need to “burst” that load count under certain conditions.
The second option is to install only a Head Node and a Broker Node onsite, and host all Compute Nodes in Windows Azure. This is often the pattern for organizations that need HPC on a scheduled and periodic basis, such as financial analysis or actuarial table calculations.
Blog entry on Hybrid HPC with Windows Azure: http://blogs.msdn.com/b/ignitionshowcase/archive/2010/12/13/high-performance-computing-on-premise-and-in-the-windows-azure-cloud.aspx
Links for further research on HPC, includes Windows Azure information: http://blogs.msdn.com/b/ncdevguy/archive/2011/02/16/handy-links-for-hpc-and-azure.aspx