Gregory Allen (The University of Texas at Austin), “Computational Process Networks: A Model and Framework for High-Throughput Signal Processing” (2011)

Gregory Allen (The University of Texas at Austin), “Computational Process Networks: A Model and Framework for High-Throughput Signal Processing” Advisor: Prof. Brian L. Evans (2011)

Many signal and image processing systems for high-throughput applications require concurrent implementations. These implementations should scale with increasing availability of computational elements, avoid deadlock, and produce consistent results (determinism). However, coarse-grained locks yield systems that do not scale well, and insufficient locking may cause non-determinate execution. In standard industry approaches, the programmer must try to resolve the tension among scalability, deadlock and determinism, which becomes increasingly difficult as software complexity grows. This dissertation presents the Computational Process Networks (CPN) model. CPN is built on a formal, mathematically provable model of concurrency, yields determinate execution, and naturally describes processing of streaming data samples, esp. sliding window algorithms. I also present a scalable implementation of the CPN model in C++ on multicore and distributed systems. Case studies include a high-performance 3-D sonar beamformer. Lastly, I develop a scalable, determinate algorithm to detect and resolve deadlocks in CPN and other process network models.

For details, please access the full thesis or contact the author.