Back-off Regulator for Improved Throughput, Congestion Avoidance and Fairness

In this thesis, we present an architectural component, the task request regulator, for flow control of incoming server-task requests that are made by a large-scale number of clients. The challenge is to keep the server at high utilization levels while avoiding overloads. Our solution is based on ad-hoc (re)scheduling of incoming client servertask requests. Namely, the regulator can order the client to back-off and return at a server-convenient time. Our solution includes a regulator that monitors the server load and tries to keep the number of client-requests at service at a preferable level. We have designed and demonstrated, both analytically and experimentally, three algorithms for implementing the regulator. The first algorithm is elegant, has modest implementation requirements but provides no fairness guarantees. The second algorithm has a shorter convergence period than the first one, at the expense of a modest increase in the storage and communication costs (but provides no fairness). Our third proposal is an extension of the first two algorithms which provides fairness with respect to the number of rescheduling events that a task may get, at a small added computational cost for the regulator.