Summary of LISA17 - Queueing Theory in Practice: Performance Modeling for the Working Engineer
Eben gives a talk on Queueing Theory in practice, specifically focusing on performance modeling for engineers. He emphasizes the benefits of Queueing Theory for builders and operators of complex systems, providing a vocabulary and toolkit for modeling software systems. The talk is divided into two main sections: Modeling Serial Systems and Modeling Parallel Systems. In the first section, Eben discusses how to build and apply a simple model for a serial system, addressing topics such as average wait time and variability in test arrivals. In the second section, he explores the performance of a fleet of servers working in concert and introduces the idea of randomized load balancing and iterative partitioning. The talk concludes by encouraging system designers to make smart compromises based on the concepts and lessons learned. Speakers: - Eben
Notable Quotes
—
01:12
— « The single most effective thing you can do to improve client observed latency but to improve capacity is to make your service faster in the first place. »
—
12:23
— « Variability is bad. If tests didn't arrive at random, if they were perfectly uniformly sized and showed up at perfectly uniform intervals then they wouldn't ever collide with each other, there would be no queuing. »
—
13:38
— « The more n increases, the more time we spend just assigning tasks, the less time we spend actually doing the work. »
—
23:06
— « Now we can keep adding servers and keep improving throughput. »