Farming out : a study.

Farming is one of severals ways of arranging for a group of individuals to perform work simultaneously. Farming is attractive. It is a simple concept, and yet it allocates work dynamically, balancing the load automatically. This gives rise to potentially great efficiency; yet the range of applications that can be farmed efficiently and which implementation strategies are the most effective has not been classified.

This research has investigated the types of application, design and implementation that farm efficiently on computer systems constructed from a network of communicating parallel processors. This research shows that all applications can be farmed and identifies those concerns that dictate efficiency. For the first generation of transputer hardware, extensive experiments have been performed using Occam, independent of any specific application. This study identified the boundary conditions that dictate which design parameters farm efficiently. These boundary conditions are expressed in a general form that is directly amenable to other architectures. The specific quantitative results are of direct use to others who wish to implement farms on this architecture.

Because of farming’s simplicity and potential for high efficiency, this work concludes that architects of parallel hardware should consider binding this paradigm into future systems so as to enable the dynamic allocation of processes to processors to take place automatically. As well as resulting in high levels of machine utilisation for all programs, this would also permanently remove the burden of allocation from the programmer.