Cloud speeds up Westpac simulations
When Westpac principal architect Ward Britton found that the bank's quantitative analysts were unable to do anything while their desktops ran financial simulations, he wondered whether he could use the opportunity to test offloading the calculations to the cloud and get a performance boost.
Speaking at Microsoft's Tech.Ed event on the Gold Coast this week, Britton described the process that the bank went through to solve its problem. The bank was using Numerix in combination with Excel 2010 to run its Monte Carlo financial simulations. Monte Carlo simulations are often used to calculate company value or evaluate investments. During the simulation process, Westpac found that its desktops were unusable. After talking with Numerix to determine the best way to reclaim its desktops, the bank settled on Windows HPC Server 2008 R2, and decided to build a proof of concept to test it.
In addition to alleviating CPU and RAM demands on its desktops, Britton said that the bank also wanted to be able to run simulations on demand, rather than as a scheduled batch job.
The bank offloaded the Excel workbooks from the desktop to an HPC head node, breaking the task up into smaller chunks and assigning them compute nodes, freeing up the computing power and RAM of the desktops.
One of the key concerns that arises when financial institutions consider the cloud was whether customer data could be stolen or leaked. However, Britton demonstrated that all of the simulations contained non-customer information.
Britton said that the bank did run into difficulties when sending the information beyond its corporate firewalls. There were several ports that the proof of concept required access to in order to operate, which, by default, were closed off to traffic. Britton said that a workaround was developed to intercept requests on specific ports and re-route them through a SOCKS proxy; however, he said that he had to work very closely with the bank's security team, and didn't recommend it as a permanent solution.
"The Azure-HPC integration that we were using for this proof of concept didn't support a proxy," he said, adding that this lack of support was a fundamental design challenge, given that most large organisations were behind firewalls and the cloud was, by nature, always on the other side of it.
"The security challenge consumed quite a lot — the lion's share — of the time of this proof of concept," he said.
However, after the quirks were ironed out, Britton said that there was a drastic drop in the time needed to run simulations. A task that would normally take over 300 minutes was completed in about 25 minutes using six compute nodes.
The time needed to run the simulations dropped as more compute nodes were used, but Britton also observed that when more than 20 to 30 compute nodes were used, the process tended to slow down. The bank isn't sure why this was the case, but it did suspect that it may be a software issue.
Britton said that problems experienced during the build of the proof of concept were also easier to fix, due to the use of a single vendor, saying that when things went wrong, there was "only one throat to choke".
Michael Lee travelled to Tech.Ed as a guest of Microsoft Australia.