Computer Science Location: Zoom
Add to Calendar 2020-06-09T10:00:00 2020-06-09T10:00:00 America/New_York PhD Proposal Presentation by Vanamala Venkataswamy Scheduling to ensure performance and cost effectiveness in data centers running on renewable energy Abstract:  Zoom

Scheduling to ensure performance and cost effectiveness in data centers running on renewable energy

Abstract: 

The majority of data centers get electricity from the electric grid: electricity generated using non-renewable sources. Electricity cost significantly contributes to the operational expense in data centers. Higher electricity costs translate to the end-users paying higher prices for the cloud services or reduced profits for the service providers. We need cheaper alternative renewable energy sources to power the data centers. Moreover, renewable energy sources significantly reduce carbon footprint. 

Using renewable energy sources to power the "green data centers" has challenges. For instance, wind farms, wind flow is not continuous across all regions, time of day, or all seasons. With green data centers, the goal is to provide carbon-free, reliable cloud services to users at a fraction of the cost of traditional cloud service providers and be able to utilize the TeraWatts of surplus zero-carbon electricity. 

Three significant issues that need addressing are 1) Meeting user Service Level Objectives (SLOs), 2) Managing distributed resources, and 3) Power variability. Hand-engineering domain specific heuristics-based schedulers to meet specific objective functions is time-consuming, expensive, and needs extensive fine tuning. My goal is to explore the possibility of applying Reinforcement Learning (RL) to learn effective job scheduling policies automatically that work in this complex dynamic environment.

I propose three job scheduler frameworks that are aware of the power variability at multiple green data centers and effectively schedule users' jobs. The objective function for our schedulers is to maximize meeting user SLOs by maximizing value specified in the SLOs (maximizing revenue for the service providers). 

The proposed schedulers leverage RL to learn workload specific scheduling policies without human inputs beyond high-level objectives while meeting user SLOs. Applying the RL framework is a novel approach to schedule jobs in a complex dynamic green data center environment.

Limited to UVA Faculty, staff, and funded graduate students.

Committee:

  • Dr. John Stankovic (Chair)
  • Dr. Andrew Grimshaw (Advisor)
  • Dr. Don Brown (External)
  • Dr. Yanjun Qi
  • Dr. Haifeng Xu