This project is lead by Dr. Vijay Narayanan at Penn State University. We are trying to build a video analytics platform. In this project, we will first focus on the in-storage acceleration of video analytics workload. For example, how we may shift some simple video processing workload to SSD controller. A common job of the SSD controller involves garbage collection. In the very first part of the project, we want to see how we may improve the efficiency of garbage collection if a host server only performs video-processing-related operations. To achieve this, we shall install ‘Scanner’ [1], an open-source video processing software, and keep track of the SSD access patterns, as well as data being written to SSDs. We hope to identify some characteristics from this project.
This project is planned into the following steps:
- We will use Scanner as a sample video analytics workload.
- We will collect SSD access patterns as well as data being written to SSD at each SSD write.
- we will analyze the collected data and optimize SSD controller garbage collection algorithm based on observations.
References:
[1] Scanner: Efficient Video Analysis at Scale. http://scanner.run/