RPM: Reward Power Manager for Power Distribution over a Cluster

The amount of power required to run an exascale system has become a challenging factor, as the recommended power budget for such systems is typically around 2030 MW. These systems must operate power-efficiently to achieve exascale computation under a limited power budget. Hardware overprovisioning is a widely used technique to address this challenge, which allows for the inclusion of more compute nodes in a cluster without exceeding the global power budget. Overprovisioning can be achieved by limiting the power allocation to each compute node below its Thermal Design Power (TDP) limit [1] . This approach is commonly known as system-wide power capping (PCAP). Modern processors now support PCAP, which enforces a user-set PCAP limit by throttling the processor’s frequency at the hardware level. However, uniform power allocation across all nodes can severely impact the performance of jobs that have high power needs. In such scenarios, non-uniform power distribution across nodes while maintaining the global power budget can prevent performance degradation and power wastage.