Analyzing Data Access Characteristics of Deep Learning Workloads and Implications

In this paper, we analyze the data access characteristics of deep learning workloads and observe that they differ significantly from traditional workloads. In memory access, access types (instruction vs. data), operations (read vs. write), access bias, and access distributions show unique patterns. In file access, a large volume of data are randomly accessed and their access bias is very weak. For this reason, it is difficult to efficiently manage memory and file data in deep learning workloads. Also, a large volume of data accessed during the training phase of deep learning can lead to severe memory thrashing. To cope with this situation, we present a new memory architecture that makes use of persistent memory to accelerate deep learning memory systems. By considering the access characteristics of deep learning, our experiments show that the proposed architecture with our preliminary management policy improves memory performance by 80% compared to the second-chance algorithm under the same persistent memory based architecture and 98% compared to an original memory architecture without persistent memory.