ETE: a customizable approach to measuring end-to-end response times and their components in distributed systems

Detecting and resolving performance problems in distributed systems often requires measurements of end-to-end (“finger tip to eyeball”) response times. Existing approaches embed transaction definitions in instrumentation codes. As a result, service providers (e.g., ISPs) cannot tailor transaction definitions to the usage patterns of their customers. We propose a new approach-ETE (end-to-end)-in which transaction definitions are externalized so that they can be customized. This is accomplished by having instrumentation generate events (not transactions) and employing a separate component-the transaction generator-that uses external definitions of transactions to construct response time measurements from event streams. ETE provides measurements of both end-to-end response times and their components. The latter reflect delays for services within distributed systems (e.g., name resolution service). We have used ETE to measure response times for Web transactions, terminal emulators, and Lotus Notes