A Case Study in Computational Caching Microservices for HPC

A case study is presented that provides computation caching (memoization) through a microservice architecture to high-performance computing (HPC) applications, particularly the ExMatEx proxy application CoEVP (Co-designed Embedded ViscoPlasticity Scale-bridging). CoEVP represents a class of multiscale physics methods in which inexpensive coarse-scale models are combined with expensive fine-scale models to simulate physical phenomena scalably across multiple time and length scales. Recently, CoEVP has employed interpolation based on previously executed fine-scale models in order to reduce the number of fine-scale evaluations needed to advance the simulation. Building on this work, we envision that distributed microservices composed to provide new capabilities to large-scale parallel applications can be an important component in simulating ever-larger systems at ever-greater fidelities. We explore three aspects of a microservice composition for interpolation-based memoization in our study. First, we present a cost assessment of CoEVP's current fine-scale modeling and interpolation approach. Second, we present an alternative interpolation strategy in which interpolation models are directly constructed on demand from previous fine-scale evaluations: a "database of points" rather than a "database of models." Third, we evaluate the characteristics of the two approaches with and without cross-process sharing of database entries. Lessons learned from the study are used to inform designs for future work in developing distributed, large-scale memoization services for HPC.