Temporal dependence within the workload of any computing or networking system has been widely recognized as a significant factor affecting performance. More specifically, autocorrelation in flows, is catastrophic for performance. In a simple single server system, autocorrelation in the arrival intensities or service demands may result in user response times that are slower by several orders of magnitude. In homogeneous clusters where size-based load balancing policies have been proved optimal for performance, autocorrelation in the arrival intensities of jobs obliterates any performance benefit of traditional load balancing policies. In multi-tiered systems, if a service process of any of the tiers is autocorrelated, then user response times are very high, in spite of the fact that the bottleneck resource in the system is far from saturation and that the measured throughput and utilizations in all other tiers are also modest, falsely indicating that the system can sustain higher capacities. In storage systems, autocorrelation in the arrival or service processes at the disk level may result in significant user-perceived performance degradation.
This project aims at providing a practical way to characterize and quantify the performance impacts of autocorrelated flows in systems. The main focus is on the development of new technologies for resource allocation that consider autocorrelation as an important characteristic of any stochastic process. On-line monitoring of autocorrelation provides the necessary information for scheduling parameterization, making an important step toward the development of autonomic systems.
