Publications

Hardware overprovisioning is widely used to cap processor power consumption and improve average power utilization across data-center servers and HPC nodes. This work introduces Themis, a power management library that characterizes the power sensitivity of co-running applications and independently adjusts core and uncore frequencies to maximize system throughput under power constraints.

Themis dynamically determines and redistributes surplus power and reward power based on application power sensitivities. Evaluation across exascale proxy application mixes and power caps demonstrates improved system throughput over existing approaches.

This paper introduces Fulcrum, a power management library for co-running parallel applications on multi-socket, multi-core servers. Fulcrum dynamically redistributes power under a global power constraint and independently adjusts core and uncore frequencies according to application power sensitivity and degree of parallelism.

After optimizing application-level power usage, Fulcrum further reallocates power to balance performance gains across applications and improve fairness.

KarmaPM is a power management library for co-running applications on multiprocessor systems. It dynamically redistributes power bidirectionally across sockets while maintaining fairness between applications.

KarmaPM profiles CPU utilization, identifies underutilizing donor sockets, transfers surplus power to high-demand receivers, and uses a reward-power scheme when donor applications return to high-utilization phases.

Cuttlefish is a programming-model-oblivious C/C++ library for achieving energy efficiency in multicore parallel programs running on Intel processors. An online profiler discovers application memory access patterns and combines DVFS and uncore frequency scaling to dynamically adapt processor frequencies.

This thesis work addresses power management for co-running applications through dynamic redistribution of surplus power and application-aware frequency selection. The proposed two-stage approach first improves throughput through bidirectional power redistribution and then balances performance gains to improve fairness.