Invited Paper: ServerFilling: A better approach to packing multiserver jobs.

Ever since the advent of "multiserver jobs" (jobs that require more than one server or core simultaneously), practitioners have been faced with the question of how to pack these jobs into a compute cluster. While many policies have been proposed, including First-Come-First-Served (FCFS), BackFilling, MaxWeight, and Most Servers First, it is not well understood which policies simultaneously achieve (1) throughput-optimality and also (2) both low and theoretically predictable mean queueing times. This paper reviews some very recent work from [9, 10] on an alternative packing policy called ServerFilling (SF) and some extensions of this policy. The SF policy achieves both goals (1) and (2) above. This paper discusses and evaluates existing policies in comparison to SF, in order to prompt discussion on the tradeoffs between different scheduling policies.