HyperLink   DeepStore: In-Storage Acceleration for Intelligent Queries
   
Publication Year:
  2019
Authors
  Vikram Sharma Mailthody, Zaid Qureshi, Weixin Liang, Ziyan Feng, Simon Garcia de Gonzalo, Youjie Li, Hubertus Franke, Jinjun Xiong, Jian Huang, Wen-mei Hwu
   
Published:
  In the Proceedings of the 52 Annual IEEE/ACM International Symposium on Microarchitecture (MICRO'19), Columbus, OH, USA
   
Abstract:

Recent advancements in deep learning techniques facilitate intelligent-query support in diverse applications, such as content-based image retrieval and audio texturing. Unlike conventional key-based queries, these intelligent queries lack efficient indexing and require complex compute operations for feature matching. To achieve high-performance intelligent querying against massive datasets, modern computing systems employ GPUs in-conjunction with solid-state drives (SSDs) for fast data access and parallel data processing. How-ever, our characterization with various intelligent-query workloads developed with deep neural networks (DNNs), shows that the storage I/O bandwidth is still the major bottleneck that contributes56%90% of the query execution time.

To this end, we present DeepStore, an in-storage accelerator architecture for intelligent queries. It consists of (1) energy-efficient in-storage accelerators designed specifically for supporting DNN-based intelligent queries, under the resource constraints in mod-ern SSD controllers; (2) a similarity-based in-storage query cache to exploit the temporal locality of user queries for further performance improvement; and (3) a lightweight in-storage runtime system working as the query engine, which provides a simple soft-ware abstraction to support different types of intelligent queries.DeepStore exploits SSD parallelisms with design space exploration for achieving the maximal energy efficiency for in-storage accelerators. We validate DeepStore design with an SSD simulator and evaluate it with a variety of vision, text, and audio based intelligent queries. Compared with the state-of-the-art GPU+SSD approach, DeepStore improves the query performance by up to 17.7×, and energy-efficiency by up to 78.6×



*Vikram Sharma Mailthody and Zaid Qureshi are co-primary authors.