Performance Modeling and Evaluation of a Production Disaggregated Memory System

High performance computers rely on large memories to cache data and improve performance. However, managing the ever-increasing number of levels in the memory hierarchy becomes increasingly difficult. The Disaggregated Memory System (DMS) architecture was introduced in recent years for better memory utilization. DMS is a global memory pool between the local memories and storage. To leverage DMS, we need a better understanding of its performance and how to exploit its full potential. In this study, we first present a DMS performance model for performance evaluation and analysis. We next conduct a thorough performance evaluation to identify application-DMS characteristics under different system configurations. Experimental tests are conducted on the RAM Area Network (RAN), a DMS implementation available at the Argonne National Laboratory, for performance evaluation. Then, the results of performance experiments are presented along with an analysis of the pros and cons of the RAN-DMS design and implementation. The counterintuitive performance results for the K-means application are analyzed at code-level to illustrate DMS performance. Finally, based on our findings, we present some discussions on future DMS design and its potential on AI applications.