Serving DNNs like Clockwork - Performance Predictability from the Bottom Up
Explore a conference talk from OSDI '20 that delves into the performance predictability of serving Deep Neural Networks (DNNs). Learn about Clockwork, a distributed model serving system designed to achieve consistent low latency for machine learning inference in interactive web applications. Discover how the researchers leverage the deterministic performance of DNN inferences to build a system that can support thousands of models while meeting strict latency targets. Examine the principles behind Clockwork's design, its ability to achieve tight request-level service-level objectives (SLOs), and its high degree of request-level performance isolation. Gain insights into addressing common-case sources of latency and curtailing tail latency caused by unpredictable execution times in model serving architectures.