Analysis of the Task Assignment based on Guessing Size (TAGS) policy

We consider the performance of Task Assignment based on Guessing Size (TAGS), a multi-host job assignment policy. This policy, which was introduced by M. Harchol-Balter, was designed to work in a non-preemptive setting with unknown job sizes. The policy which is not work preserving. It has been studied numerically, especially in the important case of Bounded Pareto job size distributions. In this paper we provide the first mathematical analysis of TAGS. We compute exact stability conditions for the policy. The stability condition is given in terms of the accuracy of Markov’s inequality as an estimator for the first moment of the job size distribution. A consequence of our analysis is that for any bounded job size distribution, TAGS cannot handle a load which is more than logarithmic in the ratio of longest to shortest job. In the case of Bounded Pareto job size distributions we show that certain approximation equations for average waiting time are conservative and relatively accurate. We then compute near optimal cutoffs in this setting. We show that at low loads the penalty for not knowing job sizes is bounded by a factor of 2. We introduce a variant of TAGS which we call T+F which is much more stable than the original version. When using T+F, the penalty for not knowing job sizes is reasonable for all but the highest loads. Finally, we compare the performance of TAGS and FIFO using actual simulations rather than approximation formulas and explain when each one is preferable.