Spencer: Interactive Heap Analysis for the Masses

Abstract: Programming language-design and run-time-implementation require detailed knowledge about the programs that users want to implement. Acquiring this knowledge is hard, and there is little tool support to effectively estimate whether a proposed tradeoff actually makes sense in the context of real world applications. Ideally, knowledge about behaviour of "typical" programs is 1) easily obtainable, 2) easily reproducible, and 3) easily sharable. We present Spencer, a web service and API framework for dynamic analysis of a continuously growing set of traces of standard program corpora. Users do not obtain traces on their own, but can instead send queries to the web service that will be executed on a set of program traces. Queries are built in terms of a set of query combinators that present a high level interface for working with trace data. Since the framework is high level, and there is a hosted collection of recorded traces, queries are easy to implement. Since the data sets are shared by the research community, results are reproducible. Since the actual queries run on one (or many) servers that provide analysis as a service, obtaining results is possible on commodity hardware. Data in Spencer is meant to be obtained once, and analysed often, making the overhead of data collection mostly irrelevant. This allows Spencer to collect more data than traditional tracing tools can afford within their performance budget. Results in Spencer are cached, making complicated analyses that build on cached primitive queries speedy.