This chapter presents an argument that the composition of the workload on an individual host is important, and suggests that by maintaining an appropriate mix of processes based on the resources that they consume, the use of system resources will be improved and the onset of bottlenecks reduced. This argument is embodied in the first and central process mix hypothesis.
There are two physical means for distributing processes to adjust the mix of processes. Migration can monitor the activity of a process and make adjustments to the workload composition at any time. Initial placement is more restrictive, as distribution occurs before any activity can be monitored. Thus policies for initial placement mechanisms require more information than is available at distribution time and must use some form of prediction. Thus the prediction accuracy and initial placement hypotheses are presented while assuming the process mix hypothesis.
The final combination hypothesis suggests that if the initial placement hypothesis fails, the combination of initial placement and process migration may solve the prediction accuracy problem.
Chapter 8 describes the algorithms developed to test these hypotheses and chapter 4 details related work.