Latency and Bandwidth Considerations in Parallel Robotics Image Processing

Abstract

Parallel image processing for robotics applications differs in a fundamental way from parallel scientific computing applications: the problem size is fixed, and latency requirements are tight. This brings Amdhal's law in effect with full force, so that message-passing latency and bandwidth severely restrict performance. The authors examine an application from this domain, stereo image processing, which has been implemented in Adapt, a niche language for parallel image processing implemented on the Carnegie Mellon-Intel Corporation iWarp. High performance has been achieved for this application. It is shown how a I/O building block approach on iWarp achieved this, and the implications of this performance for more traditional machines that do not have iWarp's rich I/O primitive set are examined.