A Real-Time Parallel Application

We deal with the detection of gravitational chirp signals among noisy data, where the reception and the detection are piped and run in parallel. We consider the classical theory of signal detection, which yields a detector with a “bank-of-filters” structure. We investigate distributed network computing in order to implement such a detector by heterogeneous high performance workstations interconnected via an Ethernet network. The goal is to design a distributed detector running on a number of available workstations. The computation is decomposed across the workstations in such a way to minimize communications and to match the acquisition rate. Our approach is general and can be used for networks of workstations different from those used in our experimentation. We point out that the classical performance analysis seems inappropriate if applied to real-time detection by heterogeneous distributed systems, because the execution time requirements are disregarded. To take into account such constraints we characterize the algorithm, evaluate performances on different workstations, and propose a task decomposition strategy assigning the appropriateGrainto each workstation.