CSE Faculty Seminar Series – Barzan Mozafari

Barzan Mozafari, Ph.D.

Morris Wellman Assistant Professor

Department of Electrical Engineering and Computer Science

University of Michigan

“A Journey From Faster to More Predictable: Using Statistics to Build Better Data-Intensive Systems“

Abstract: With much of the database research focused on improving raw performance (throughput, mean latency), performance predictability has often been neglected. The large gap between mean and high percentile latencies of modern systems is a major cause of hardware over-provisioning and expensive redundancies in data and computation.  In this talk, we begin a journey by vetting various notions of predictability that affect data-intensive applications. We then focus on the internal sources of performance unpredictability and introduce a new tool, called VProfiler. Unlike traditional profilers, VProfiler pin points the root causes of performance variance in a complex codebase. Our findings lead us to a new scheduling algorithm, which has now become the default policy in MySQL. We then turn our attention to external sources of unpredictability, and study the applicability of Robust Optimization theory. Our journey from faster to more predictable comes full circle, as we discover that speed and predictability are not mutually exclusive; predictable performance is often faster too! We conclude by introducing a practical tool for predicting future performance based on past statistics.

Bio: Barzan Mozafari leads a research group that designs the next generation of scalable databases using statistical techniques. His research career has led to several open-source projects, including DBSeer, BlinkDB, and SnappyData. He is the recipient of an NSF CAREER award and several best paper awards. 

Lunch will be provided