A processor allocation framework that uses Amdahl’s Law to apportion resources fairly among data center users has won the best paper award at the 2018 International Symposium on High-Performance Computer Architecture (HPCA). Seyed Majid Zahedi and Qiuyun Llull, co-first authors on the paper, worked with Professor Benjamin Lee, Associate Professor in the Departments of Electrical and Computer Engineering and Computer Science, to develop a framework that guarantees fairness yet outperforms existing mechanisms for allocating resources. Zahedi is a graduate student in Computer Science and will receive his doctorate this year. Llull recently completed her doctorate in Electrical and Computer Engineering and is now a performance engineer at VMware.
When users share large-scale computer systems, they seek guarantees on their resource allocations and determine whether pooling their resources into a large, shared system is beneficial. To produce these guarantees, the team drew on Amdahl’s Law. This law from computer architecture, first presented in 1967, models performance as a job is allocated processor cores. Inspired by Amdahl’s Law, the team proposed the Amdahl utility function to model users’ utility from processor core allocations and designed the Amdahl bidding mechanism to find allocations that optimize utilities. The bids produce a market equilibrium that provides users their expected resources.
Prof. Lee’s research group has been building interdisciplinary bridges between computer systems management and economic game theory for several years. The HPCA best paper award from 2018 follows another best paper award from the International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) from 2016, which recognized co-first authors Songchun Fan and Seyed Majid Zahedi for their research on using game theory to manage shared data center power. Fan recently completed her doctorate in Computer Science at Duke and is now a software engineer at Google.
Editor's note: This revised article represents the contributions of all research participants more accurately.
From left: Songchun Fan, Qiuyun Llull, Ben Lee, and Seyed Zahedi