Ancient and new algorithms for load balancing in the ℓ_p norm

We consider the on-line load balancing problem where there are m identical machines (servers) and a sequence of jobs. The jobs arrive one by one and should be assigned to one of the machines in an on-line fashion. The goal is to minimize the sum (over all machines) of the squares of the loads, instead of the traditional maximum load. We show that for the sum of the squares the greedy algorithm performs within 4/3 of the optimum, and no on-line algorithm achieves a better competitive ratio. Interestingly, we show that the performance of Greedy is not monotone in the number of machines. More specifically, the competitive ratio is 4/3 for any number of machines divisible by 3 but strictly less than 4/3 in all the other cases (although it approaches 4/3 for a large number of machines). To prove that Greedy is optimal, we show a lower bound of 4/3 for any algorithm for three machines. Surprisingly, we provide a new on-line algorithm that performs within 4/3 - δ of the optimum, for some fixed δ > 0, for any sufficiently large number of machines. This implies that the asymptotic competitive ratio of our new algorithm is strictly better than the competitive ratio of any possible on-line algorithm. Such phenomena is not known to occur for the classic maximum load problem. Minimizing the sum of the squares is equivalent to minimizing the load vector with respect to the ℓ₂ norm. We extend our techniques and analyze the exact competitive ratio of Greedy with respect to the ℓ_p norm. This ratio turns out to be 2 - Θ((In p)/p). We show that Greedy is optimal for two machines but design an algorithm whose asymptotic competitive ratio is better than the ratio of Greedy.