Near Data Processing Performance Improvement Prediction via Metric-Based Workload Classification

Abstract

Contrary to the improvement of CPU capabilities, traditional DRAM evolution faced significant challenges that render it the main performance bottleneck in contemporary systems. Data-Intensive applications such as Machine Learning and Graph Processing algorithms depend on time and energy consuming transactions between the memory bus and the CPU caches. The emergence of 3D-Stacked memories that provide a very high bandwidth led to the exploration of the Process-In-Memory (PIM) paradigm where logic is added to the memory die and data are being processed where they reside. To fully exploit this model, there is a need to methodically determine the portions of code that are better fitted for Near-Data-Processing (NDP). To this extend, in this work, after presenting the key trends of the research field and examine proposed criteria, we simplify the process of a priori decision of a block’s suitability by proposing a two-step metric-based application categorization able to predict the applications behavior when offloaded for NDP.