Performance Analysis of Cloud Applications

Many popular cloud applications are large-scale distributed systems with each request involving tens to thousands of RPCs and large code bases. Because of their scale, performance optimizations without actionable supporting data are likely to be ineffective: they will add complexity to an already complex system often without chance of a benefit. This paper describes the challenges in collecting actionable data for Gmail, a service with more than 1 billion active accounts.

Using production data from Gmail we show that both the load and the nature of the load changes continuously. This makes Gmail performance difficult to model with a synthetic test and difficult to analyze in production. We describe two techniques for collecting actionable data from a production system. First, coordinated bursty tracing allows us to capture bursts of events across all layers of our stack simultaneously. Second, vertical context injection enables us combine high-level events with low-level events in a holistic trace without requiring us to explicitly propagate this information across our software stack.