William Knebel, Marc R. Gastonguay, Dan Polhamus, and Jeffrey T. Hane
Metrum Research Group, LLC, Tariffville, CT, USA
Objectives: In theory, elastic cloud computing (EC2) could virtually remove computing time from the critical path of pharmacometric (PM) projects. Optimization of the relationships between available EC2 resources, modeling strategies, and scientist teams, will be required to achieve the full performance potential. The objectives of this work were to capture the usage patterns for a typical group of modeling and simulation (M&S) scientists, and to summarize strategies for effective use of EC2 resources for PM analyses.
Methods: Usage patterns, based on the number of simultaneous compute cores (SCC) utilized, for a group of M&S scientists with access to Amazon EC2 based virtual machines, were captured to assess sustained and peak use of cloud based resources over a 6 month period. SCC usage reflected a combination of NONMEM(R), R, and/or OpenBUGS software. Actual usage patterns for a group of 14 individual users were recorded. These data were used to simulate via resampling, the usage patterns for groups of up to 64 users. A strategy for efficient EC2 resource utilization was also illustrated by summarizing an actual PM project. This case-study included population model development for 3 endpoints requiring numerical integration of differential equations and simulation-based evaluation of Phase 2 and 3 trial designs options.
Results: The most active user over the 6 month period demonstrated multiple usage peaks of 150 SCC, with several days of less than 10 SCC. For the group of 14 users, peak usage of 750 SCC was demonstrated on a single day with sustained usage of 250 SCC over the 6 month period. Simulation of the group of 64 users revealed sustained usage of 1500 SCC with intermittent peaks of 2400 SCC over 6 months. For the case study, a strategy of parallel task and computation implementation across a team of 6 M&S scientists, and the EC2 infrastructure, allowed project completion within 14 days, where linear sequential task implementation with a single scientist and EC2 would have required 8 weeks, and linear sequential implementation without EC2 would have required 4 months or more.
Conclusions: Usage of EC2 resources, and associated software, was proportional to the number of users, not maximum number of available cores. Team-based project strategies, with parallel task and computation implementation, maximize the potential utility of EC2 for PM workflows.
Reference: PAGE 23 (2014) Abstr 3136 [www.page-meeting.org/?abstract=3136]
Poster: Methodology - Other topics