Today's production clusters host a variety of jobs with diverse resource demands, and fair allocation of the cluster resources is critical for performance isolation among those jobs. Nonetheless, existing resource allocation policies implicitly assume that a job can run at exactly the same efficiency with any unit of its usable resources, which usually does not hold in practice. In fact, we find in many scenarios that a job can have different preferences for different resources. For example, a job can run much faster in those machines directly storing its input data, although it can still run in other ones after paying a performance penalty. Such heterogeneous resource preferences are called by us as soft-constraints, and it is still unclear how to define and achieve fairness in...[ Read more ]

Today's production clusters host a variety of jobs with diverse resource demands, and fair allocation of the cluster resources is critical for performance isolation among those jobs. Nonetheless, existing resource allocation policies implicitly assume that a job can run at exactly the same efficiency with any unit of its usable resources, which usually does not hold in practice. In fact, we find in many scenarios that a job can have different preferences for different resources. For example, a job can run much faster in those machines directly storing its input data, although it can still run in other ones after paying a performance penalty. Such heterogeneous resource preferences are called by us as soft-constraints, and it is still unclear how to define and achieve fairness in the presence of soft-constraints.

In this work, we propose MTTC, a proactive exchange-based sharing policy for fair allocation with soft constraints, and show that it is the only policy satisfying all the typical fairness criteria, including an important property that prevents selfish users from lying to benefit themselves. Furthermore, we approximate the MTTC allocation by an online preference-aware scheduler called FSC, and have integrated the FSC prototype into Apache YARN. The effectiveness of FSC is confirmed with both testbed experiments in a 65-node Amazon EMR cluster and trace-driven simulations. Particularly, the simulation results suggest that FSC can reduce the average job completion time by over 54%.