| Title | : | Curbing Energy Wastage In Data Centers: A Systems Approach |
| Speaker | : | Kanad Ghose (SUNY Binghamton) |
| Details | : | Mon, 4 Feb, 2019 4:00 PM @ A M Turing Hall |
| Abstract: | : | Data centers are designed to handle a peak workload volume that is not or rarely encountered in their operating lifetime. In fact, industry-wide utilization of data centers are reported to be in the range 18% to 30%. Data center operators also tend to idle servers to avoid performance penalties when load surges occur - this overprovisioing results in a tremendous energy wastage. This talk introduces Dynamic Capacity Provisioner (DCP): an energy-aware data center scheduling and management system that provisions server capacity to match the instantaneous workload. To provide better energy proportionality and energy savings, DCP concentrates work onto the smallest subset of active servers possible. Tail latency increases compared to a fully-provisioned system are limited using short-term load prediction. As workload volume decreases, servers are offlined and enter a quiescent state after they complete their previously assigned tasks and the remaining servers remain in the deep sleep states, or are powered off altogether to save energy. DCP is also thermally-aware and keeps the core temperature of active servers within specified limits while reducing thermal stresses caused by transitions between active and quiescent states. In our prototype implementation with Linux servers from multiple generations, DCP achieves a 10% to 30% reduction in overall server energy consumption on average across different system configurations and across several inherently different, bursty, and diverse workloads with minimal degradation in both response time and throughput. We also present extensions of DCP to avoid imbalances in the power drawn on the three electrical phases for an AC-powered data center that avoids further inefficiencies in the power distribution system.
This talk concludes with an overview of my current projects in other areas - the design and implementation of security extensions to an existing microarchitecture (a DARPA-funded joint effort with industrial partners) and a smart ECG patch on a flexible substrate that features onboard processing for removing noise and motion artifacts and calculating hear rate and heart rate variability parameters. Speaker Bio : Kanad Ghose (PhD, Iowa State) is a Professor in the Computer Science Department at the State University of New York at Binghamton. He served as the department's chair from 1998 to 2016. Prof. Ghose's research interests are in all aspects of energy-aware computing, processor microarchitecture and computer systems software. He serves as the site director for the Center for Energy-Smart Electronics Systems (ES2), a multi-university NSF Industry-University Collaborative Research Center led by Binghamton. He has extensive publications in his research areas, graduated numerous PhD and MS students and holds 24 US patents. His research sponsors include Federal agencies like NSF, DARPA, AFRL and the industry (Intel, Microsoft, IBM, HP, Xilinx and many others). |
