Beanstalks: A General Purpose Job Queue for the Computer Vision Cluster

The computer vision group at Berkeley EECS has a set of computers for research use, yet lacks an efficient framework to utilize the computational power. In this project, we propose a novel job queue system that is designed for the generalpurpose use in the group. Specific needs of the group such as being easy to use, multiple language support, and tolerance of different user styles are considered and reflected in the system design. We also discuss typical crash recovery and load balancing issues, and proposed ways to implement them. Finally, we discuss the experience and lessons learned during the CVPR paper deadline, and propose several ways to further improve the current system.