Addressing the Capacity Challenge through High-Performance Computing

New regulations and rapidly changing market conditions are increasing the demand for more sophisticated analysis in shorter timeframes. Van Beach, MG-ALFA Director of Business Development at Milliman explains how HPC helps actuaries and risk managers work more efficiently and effectively to handle the increased workloads.

WFS: What challenges do your clients face in today’s environment?

VB: Financial modeling has become increasingly critical given the current and pending regulatory changes and recent market volatility. Insurers face a multi-faceted capacity challenge as financial modeling requirements expand across several dimensions. Models are becoming more complex to reflect key risk characteristics. Advanced techniques require stochastic and often nested stochastic projections. More models are required and more users are involved. Finally, reporting windows are shorter, more iterations are required, and information is needed in near real-time.  

WFS: How are they addressing these challenges?

VB: They need to enable fast runtimes, complex models, efficient execution and wide user access to meet the capacity challenge. Software, infrastructure, and process advances are all needed to realize the efficient, effective modeling environment of the future.

WFS: How does HPC help insurers meet this capacity challenge?

VB: The most fundamental aspect of the capacity challenge is speed, which HPC helps address by enabling a cluster of computing resources to be utilized to complete an MG-ALFA modeling job. However, not all HPC integrations are equal. MG-ALFA has been optimized for distributing projections using HPC. As computing resources are added to the cluster, clients can realize a nearly linear reduction in runtime. We’ve proven this linearity in independent benchmark tests, and we can offer clients this incredibly efficient scalability with HPC. 

WFS: How have you addressed other aspects of the capacity challenge?

VB: When running large models with MG-ALFA, clients rarely encounter the memory constraints that often plague other systems. Recent enhancements to MG-ALFA’s memory use have enabled clients to run 200,000 model points in a full ALM projection with dependent assets and liabilities with just 2 GB of memory on a standard 32-bit machine. The model size limitation has been effectively eliminated without the need to invest in additional memory, updated hardware or new operating systems.  

HPC widens access to additional computing resources for financial modeling. The seamless integration of HPC with MG-ALFA allows new users of the cluster to be up, running and more productive quickly. Further, there is little overhead for users to submit MG-ALFA modeling jobs to the cluster using HPC.  

WFS: How do you see HPC evolving in the future?

VB: With the first deployments of HPC with MG-ALFA, clients addressed an “embarrassingly parallel” stochastic process. For example, 1,000 scenarios were distributed to a limited number of computing resources (typically discrete cores), so there were always more tasks (scenarios) than computing resources. Today those 1,000 scenarios might be submitted to 500 quad-core machines, so there are 2,000 computing resources available. The embarrassingly parallel process no longer exists as there are more computing resources than scenarios. The combination of MG-ALFA’s scalability and the cost effectiveness of HPC have created a new challenge. Milliman is working with Microsoft to define and manage tasks at a more granular level than a scenario for MG-ALFA to continue achieving near linear reductions in runtime as HPC computing resources are added. 

Modeling processes also are becoming more complex. Consider the MCEV reporting requirements proposed by the CFO Forum for analysis of movement. This report necessitates multiple sequential runs of complex MCEV models to capture the impact of incremental changes. These multi-stage modeling requirements suggest the capacity challenge will evolve beyond individual models to include managing broader modeling processes. HPC’s speed and efficiency is critical as processes grow, and Milliman will provide new solutions as these new paradigms emerge. 

www.milliman.com