Improved Algorithms for Online Load Balancing

We consider an online load balancing problem and its extensions in the framework of repeated games. On each round, player chooses a distribution (task allocation) over K servers, then environment reveals server, which determines computation time server for processing task assigned. After all rounds, cost is measured by some norm cumulative computation-time vector. The makespan if \(L_\infty \)-norm. goal to minimize regret, i.e., minimizing player’s relative best fixed hindsight. propose algorithms general norms prove their regret bounds. In particular, \)-norm, our bound matches known proposed algorithm runs polynomial per trial involving linear programming second order programming, whereas no was previously achieve bound.