Analyzing Memory Access Traces of Deep Learning Workloads for Efficient Memory Management

Due to the recent advances in artificial intelligence (AI) and high-performance computing technologies, deep learning is actively used in various application domains of the fourth industrial revolution. Since the data size of deep learning increasingly grows, analyzing the memory reference characteristics of AI workloads becomes important. In this article, we analyze the memory access traces of deep learning workloads consisting of text and image data. Specifically, we analyze read and write operations separately for instruction and data memory accesses. Based on our analysis, we find out the following characteristics that are quite different to memory access traces of conventional desktop workloads. First, when comparing instruction and data accesses, instruction access accounts for a little portion of memory accesses in deep learning workloads, which is quite different from traditional workloads. Specifically, instruction access accounts for 1-3.3% in deep learning workloads while 15-30% in traditional workloads. Second, when comparing read and write accesses, write access accounts for 64-80%, which is also different to traditional workloads where write accounts for 6-55%. Third, although write access accounts for the majority of memory accesses, it exhibits the low access skewness. Specifically, the Zipf parameter of write access is about 0.3, and the parameter is smaller in text workloads than image workloads. Fourth, in predicting re-reference likelihood, recency ranking is important in instruction and data read accesses, but frequency ranking is necessary in write accesses for accurate estimation. We expect that the characterization study in this article will guide for managing memory systems of deep learning workloads.