Authors

Date

2006

Scope

Language

Identifiers

- handle: 10670/1.4vzmq1
- hal: hal-00341356
- doi: 10.1007/11823285

Organization

Eric Angel et al., « The price of approximate stability for scheduling tasks on two links », *Hyper Article en Ligne - Sciences de l'Homme et de la Société*, ID : 10.1007/11823285

We consider a \em scheduling game, where a set of selfish agents (traffic loads) want to be routed in exactly one of the two parallel links of a system. Every agent aims to minimize \em her own completion time, while the social objective is the \em makespan, i.e. the time at which the last agent finishes her execution. We study the problem of optimizing the makespan under the constraint that the obtained schedule is a (pure) Nash equilibrium, i.e. a schedule in which no agent has incentive to unilaterally change her strategy (link). We consider a relaxation of the notion of \em equilibrium by considering $\alpha$-approximate Nash equilibria, in which agents do not have \em sufficient incentive to unilaterally change their strategies: in an $\alpha$-approximate Nash equilibrium, no agent can decrease its cost by more than $\alpha$ multiplicative factor. Our main contribution is the study of the \em tradeoff between the \em approximation ratio for the makespan and the stability of the schedule (the value of $\alpha$). We first give an algorithm which provides a solution with an approximation ratio of $\frac87$ for the makespan and which is a 3-approximate Nash equilibrium, provided that the local policy of each link is \em Longest Processing Time (LPT). Furthermore, we show that a slight modification of the classical \em Polynomial Time Approximation Scheme (PTAS) of Graham allows to obtain a schedule whose makespan is arbitrarily close to the optimum while keeping a constant value for $\alpha$. Finally, we give bounds establishing relations between the value of $\alpha$ and the best possible value of the approximation ratio, provided that the local policies of the links are LPT. We also briefly examine the case where the local policies are SPT.