Evaluating the Run-Time Performance of Kahn Process Network Implementation Techniques on Shared-Memory Multiprocessors

Software development tools have not adapted to the growing popularity of  multi-core CPUs, and developers are still "stuck"' with low-level and  high-cost thread abstractions.  The situation is becoming even more  complicated with the advent of heterogenuous computing.  In this article, we  point out some drawbacks of high-level abstractions currently in use, and  propose Kahn process networks (KPN) as a more high-level and efficient  abstraction for developing parallel applications.  We show that the native  POSIX mechanisms (threads and message queues) perform suboptimally as an  implementation vehicle for KPNs, and we present an implementation of a  run-time environment that can execute KPNs with less overhead.  Our  evaluation shows the advantages and disadvantages of statically mapping Kahn  processes to CPUs.