Machine Learning Optimization

As increasing demands are being made on existing hardware architectures and the cost of Persistent Memory (PM) decreases, changing how tiers of memory are organized might allow for increases in program performance with only marginal increases in systems cost. In particular, the use of Hybrid Memory architectures incorporating both Dynamic Random Access Memory (DRAM) and PM in the same tier of memory shows promise in increasing the effective RAM memory available. In order to ensure no drop in performance, however, careful management of memory allocation must occur. To this end, we have used Ruminant, a heuristic for condensing the state of Hybrid Memory into a handful of variables to allow Reinforcement Learning (RL) agents to process the state of memory and adjust the location of memory pages to best optimize allocation. This project looks to set up a learning environment to test out the performance of various RL agents on differing memory workloads and make recommendations on efficient allocation strategies.