Reducing Tail Latency of DNN-based Recommender Systems using In-storage Processing

Abstract

Most recommender systems are designed to comply with service level agreement (SLA) because prompt response to users' requests is the most important factor that decides the quality of service. Existing recommender systems, however, seriously suffer from long tail latency when the embedding tables cannot be entirely loaded in the main memory. In this paper, we propose a new SSD architecture called EMB-SSD, which mitigates the tail latency problem of recommender systems by leveraging in-storage processing. By offloading the data-intensive parts of the recommendation algorithm into an SSD, EMB-SSD not only reduces the data traffic between the host and the SSD, but also lowers software overheads caused by deep I/O stacks. Results show that EMB-SSD exhibits 47% and 25% shorter 99th percentile latency and average latency, respectively, over existing systems. © 2020 ACM.