CANCELLED: CSE Faculty Seminar: Barzan Mozafari, PhD, University of Michigan

Computer Science and Engineering Faculty Seminar Series

Barzan Mozafari, Ph.D.

Assistant Professor, Computer Science and Engineering

College of Engineering, University of Michigan

Title: The Journey From Faster to More Predictable: or How Statistics Lead to Better Data-Intensive Systems Abstract: Over the past four decades, research in the database community has had one agenda: making data-intensive applications faster. This race for faster performance has understandably deflected attention from the predictability of these systems. Modern database management systems (DBMSs) have become massive code bases with many sophisticated and interconnected components, resulting in erratic and highly unpredictable performance. To mitigate tail latencies in mission-critical applications, expensive redundancies in data and computation are employed. In many cases, hardware resources are also overprovisioned by 10-30 times the peak load, and highly qualified personnel are constantly occupied with performance tuning of their database applications. The unpredictable nature of these systems has thus significantly increased the total cost of ownership (TCO) of database technology, posing serious challenges to vendors, administrators, and application developers. Therefore, it is imperative that we initiate the conversation in understanding, measuring, and improving the predictability of DBMSs. In this discussion, we start our journey by vetting existing systems against different notions of predictability. First, we explore the power of statistics in returning faster, but also approximate, answers that can meet tight deadlines. We then turn to the fundamental question of how “faster” can be sacrificed for “more predictable” despite internal and external variations. Our findings naturally lead us to Robust Optimization theory. Interestingly, our journey comes full circle where we discover that speed and predictability are not mutually exclusive; in many cases, predictable performance can also be fast! Finally, we conclude by presenting a practical tool to predict what we call “workload intelligence” using past statistics.   Lunch will be provided.