Large scale 3D reconstruction, such as city’s 3D reconstruction requires a large amount of computing resources. Its running time ranges from weeks to months on a typical consumer level machine, depending on the size of dataset. In order to accelerate this process, we design and implement ZRPC, a distributed computing framework. With ZRPC, we can easily port the existing single machine 3D reconstruction program to a cluster without modifying its source code. A 3D reconstruction project can be split to small jobs. ZRPC manages these jobs dependencies, schedules them to appropriate machines, speculates straggler jobs, and restarts failed jobs. Moreover, ZRPC is a cluster resource management system. It connects machines together, abstracts CPU, GPU, memory storage, and other computing resou...[ Read more ]

Large scale 3D reconstruction, such as city’s 3D reconstruction requires a large amount of computing resources. Its running time ranges from weeks to months on a typical consumer level machine, depending on the size of dataset. In order to accelerate this process, we design and implement ZRPC, a distributed computing framework. With ZRPC, we can easily port the existing single machine 3D reconstruction program to a cluster without modifying its source code. A 3D reconstruction project can be split to small jobs. ZRPC manages these jobs dependencies, schedules them to appropriate machines, speculates straggler jobs, and restarts failed jobs. Moreover, ZRPC is a cluster resource management system. It connects machines together, abstracts CPU, GPU, memory storage, and other computing resources away from machines (physical or virtual), enabling fault-tolerant and elastic distributed systems to easily be built and run effectively. ZRPC provides a Web UI to monitor the cluster status and job status so that we can add, delete, restart jobs directly from web page. We have put ZRPC in production use and the results show that ZRPC can successfully accelerate our 3D reconstruction project.