ISVABI: In-Storage Video Analytics Engine with Block Interface

The wide use of cameras in the past decade has increased the need to process video data significantly. Due to the large volume of video data, analyzing videos to extract useful information has become a critical challenge. Several prior works have tried to accelerate video analytics workloads by offloading some operations to embedded processors within storage devices. However, most of theseworks require changing the block I/O interface of a normal solid-state drive (SSD) to a key-value interface, which in turn changes data structures within SSDs and requires re-programming existing applications when deploying these designs to existing warehouse-level data centers. In addition, when processing large videos, key-value SSDs perform two times slower than block I/O interface SSDs. In this work, we propose ISVABI, a software/firmware approach that maintains the SSD block I/O interface which provides offloaded operations for user-space video analytics workloads without requiring SSD hardware modification. We implement ISVABI on the Cosmos+ OpenSSD platform and show that the proposed ISVABI outperforms normal SSDs by 4.18x for various types of video operations while consuming 16% less power. We evaluate ISVABI on five real-world video analytics workloads and show a 1.89x end-to-end latency improvement.