Coauthors: Vinicius Gama Pinheiro and Denis Trystram

A typical high performance computing center is shared by many users who compete for the common infrastructure. The scheduler must not only ensure efficiency of the whole system - but also be fair to individual users, especially when the amount of work submitted by different users differs significantly. We present a scheduling algorithm that guarantees each user an upper bound on jobs' maximum stretch. The bound is proportional to the user's load and to the number of active users in the system - thus the stretch does not depend on the total load in the system. The algorithm, rather than on individual jobs, operates on campaigns of jobs. Each campaign consists of several jobs submitted at the same time by the same user; all the jobs from a campaign must complete before the next campaign starts. Scheduling policies in HPC centers rely on various ad hoc mechanisms to ensure fairness. Our work suggests an alternative approach with guaranteed performance.