PNNL

Abstract

Memory disaggregation is an approach to decouple compute and memory to minimize the total cost of ownership. However, analytical methods to study the impact of this approach are not readily available for high performance computing use cases. In this position paper, we propose "isoefficiency'" as an approach to analytically demonstrate the classes of algorithms that would benefit from disaggregated memory technologies as we scale to a larger number of processors. Isoefficiency of an algorithm is given by a function $N(p)$ that measures the degree to which the problem size needs to increase with $p$ (number of processors) to maintain a constant efficiency. We evaluate the isoefficiency on a CXL-based disaggregated system using sparse general matrix–matrix multiplication (SpGEMM) as the algorithm and using a 2-socket shared-memory system as the reference.