Directing the Datacenter with Machine Learning

Опубликовано 6 сентября 2016, 17:56

At the RAD Lab we are prototyping forward-looking datacenter software architectures using a three-pillar approach. The first pillar is exploiting application frameworks and languages optimized for high programmer productivity such as Ruby on Rails. Second is the deployment of machine learning to identify performance and scalability bottlenecks, create dynamic models for predicting performance, and mining runtime telemetry as well as console logs to identify operational problems; a framework we call the Director provides a closed observe/analyze/act loop into which these algorithms can be inserted. Third is a new persistent storage abstraction, SCADS (Scalable Consistency-Adjustable Data Store) designed specifically for the needs of datacenter-scale interactive applications, exposing consistency tradeoffs explicitly to the application developer in the context of an object-graph storage model deliberately similar to that provided by Rails' ActiveRecord.