Fair Load-Balancing on Parallel Systems for QoS

Many of the load-balancing algorithms used in parallel systems do not have a concern about response times: tasks (or requests) are simply dispatched to a server, which provides no guarantees about their execution times. When there is a maximum acceptable response time (i.e. deadline) for tasks to be executed, the consequences caused by the adoption of traditional algorithms for load-balancing can be catastrophic: when the system is under heavy loads, a huge amount of tasks miss their deadlines, even the faster ones. Also, the number of longer tasks that ends is very small - near to zero in all cases. In this paper we discuss why the traditional algorithms fail to provide the intended QoS capacity. Then, we present a new algorithm, "On-demand Restriction for Big Tasks (ORBITA)", which is proved, by simulation, to be a fair alternative for stressed systems since tasks of all durations have a chance to complete their execution before their deadlines are reached.