Periodic load balancing

Multiprocessor load balancing aims to improve performance by moving jobs from highly loaded processors to more lightly loaded processors. Some schemes allow only migration of new jobs upon arrival, while other schemes allow migration of jobs in progress. A difficulty with all these schemes, however, is that they require continuously maintaining detailed state information. In this paper we consider the alternative of periodic load balancing, in which the loads are balanced only at each T time units for some appropriate T. With periodic load balancing, state information is only needed at the balancing times. Moreover, it is often possible to use slightly stale information collected during the interval between balancing times. In this paper we study the performance of periodic load balancing. We consider multiple queues in parallel with unlimited waiting space to which jobs come either in separate independent streams or by assignment (either random or cyclic) from a single stream. Resource sharing is achieved by periodically redistributing the jobs or the work in the system among the queues. The performance of these systems of queues coupled by periodic load balancing depends on the transient behavior of a single queue. We focus on useful approximations obtained by considering a large number of homogeneous queues and a heavy load. When the number of queues is sufficiently large, the number of jobs or quantity of work at each queue immediately after redistribution tends to evolve deterministically, by the law of large numbers. The steady-state (limiting) value of this deterministic sequence is obtained as the solution of a fixed point equation, where the initial value is equal to the expected transient value over the interval between successive redistributions conditional on the initial value. A refined approximation based on the central limit theorem is a normal distribution, where the mean and variance are obtained by solving a pair of fixed-point equations. With higher loads, which is natural to consider when load balancing is performed, a heavy-traffic limit theorem shows that one-dimensional reflected Brownian motion can be used to approximately describe system performance, even with general arrival and service processes. With these approximations, we show how performance depends on the assumed arrival pattern of jobs and the model parameters. We do numerical calculations and conduct simulation experiments to show the accuracy of the approximations.