Ward Britton, principal architect at the bank explained that the bank carried out a proof of concept with the aid of Micrsoft, Numerix (the developer of the Numerix XL software used to perform Monte Carlo simulation in Excel) and Avanade (which provided project management and infrastructure specialists) to see if Windows HPC Server 2008 R2 could provide an answer. The objective was to "make it run way faster, make it run more reliably," he told attendees at Microsoft's Tech.Ed conference.
Part of the problem was that the calculations weren't performed every day, but when they were it was important to get the results quickly. Offloading the calculations to the HPC cluster proved successful, but the cost of on-premise hardware was an issue.
Given that no customer data is involved in this application, consideration was given to moving the calculation onto a cloud service. Cloudbursting to Azure looked easy, and the mechanics of building virtual machine images and uploading them to Azure was straightforward.
The problem was that Azure/HPC Server integration uses 'nonstandard' ports and does not make provision for using a proxy. That was overcome by using ProxyCap (a piece of software that reroutes traffic on certain ports through a SOCKS proxy), but Mr Britton noted that the bank's security team was closely involved in that part of the project.
So what was the performance improvement? See page 2.
Mr Britton said the use of an all-Microsoft environment provided seamless end-to-end integration, with a single API and management capabilities. And "when things didn't go to plan, there was only one throat to choke," he said.
Before Azure cloudbursting could be put into production, several things were needed, he said, including firewall and proxy support (a Microsoft staffer noted that HPC Server SP3, due in November, eliminates the nonstandard port issues), a non-beta version of the VM Role software for Azure, enhqncements to the Numerix XL software to support the HPC Server/Azure framework, a security due dilligence effort, and the development of operational guidelines (eg, it's important to ensure that Azure instances aren't left running when unused "otherwise the meter keeps running.")
There are also some licensing issues regarding the use of Excel on Azure, he added,
But the project did show the feasibility of cloudbursting computational loads onto Azure.
Disclosure: The writer attended Tech.Ed as the guest of Microsoft.