Sethi, Udhav2021-09-202021-09-202021-09-202021-08-25http://hdl.handle.net/10012/17426Servers are typically sized to accommodate peak loads, but in practice, they remain under-utilized for much of the time. During periods of low load, there is an opportunity to save power by quickly adjusting processor performance to match the load. Many systems do this by using Dynamic Voltage and Frequency Scaling (DVFS) to adjust the processor’s execution frequency. In transactional database systems, workload-aware approaches running in the DBMS have proved to be able to manage DVFS more effectively than the underlying operating system, as they have more information about the workload and more control over the workload. In this thesis, we ask whether databases can learn to manage DVFS effectively by observing the effects of DVFS on their workload. We present an approach that uses reinforcement learning (RL) to learn in-DBMS frequency governors. Our results show that governors learned using our technique are competitive with state-of-the-art methods, and are able to adapt to a variety of workload conditions. We also show that our method has an added advantage - it allows flexibility in tuning frequency governance to balance a power-performance trade-off. Finally, we discuss the challenges associated with using RL in this setting due to the overheads of using a learned frequency governor.endata systemsenergy efficiencylatency-criticalmachine learningreinforcement learningtransactionsdatabasesOLTPperformancelatencyenergy-awaredeep learningpolicy gradientREINFORCEMaster Thesis