We revisit from a fairness point of view the problem of online load balancing in the restricted assignment model and the 1‐∞ model. We consider both a job‐centric and a machine‐centric view of fairness, as proposed by Goel et al. (2005). These notions are equivalent to the approximate notion of prefix competitiveness proposed by Kleinberg et al. (2001), as well as to the notion of approximate majorization, and they generalize the well studied notion of max‐min fairness. We resolve a question posed by Goel et al. (2005) proving that the greedy strategy is globally O(log m)‐fair, where m denotes the number of machines. This result improves upon the analysis of Goel et al. (2005) who showed that the greedy strategy is globally O(log n)‐fair, where n is the number of jobs. Typically, n≫m, and therefore our improvement is significant. Our proof matches the known lower bound for the problem with respect to the measure of global fairness. The improved bound is obtained by analyzing, in a more accurate way, the more general restricted assignment model studied previously in Azar et al. (1995). We provide an alternative bound which is not worse than the bounds of Azar et al. (1995), and it is strictly better in many cases. The bound we prove is, in fact, much more general and it bounds the load on any prefix of most loaded machines. As a corollary from this more general bound we find that the greedy algorithm results in an assignment that is globally O(log m)‐balanced. The last result generalizes the previous result of Goel et al. (2005) who proved that the greedy algorithm yields an assignment that is globally O(log m)‐balanced for the 1‐∞ model.
