Achieving Safety for Power Shifting in Overprovisioned High Performance Computing Systems

摘要

Future large-scale parallel computer systems, including exascale systems, will operate under an overall power bound. This bound will be enforced by having nodes adhere to local power caps, the sum of which cannot exceed the global power bound. To make efficient use of the available power, protocols have been developed to shift power between nodes, for example to compensate for load imbalance or system noise. The distributed power shifting protocol needs to provide a guarantee of safety that the global power bound will never be exceeded and it needs to have acceptably low overhead under normal operating conditions. We present a variant of the Paxos protocol that guarantees safety. We give an upper bound for the communication overhead of this protocol, along with suggestions for how the overhead can be lowered.