The runtime structure of object ownership

Abstract

Object-oriented programs often require large heaps to run properly or meet performance goals. They use high-overhead collections, bulky data models, and large caches. Discovering this is quite challenging. Manual browsing and flat summaries do not scale to complex graphs with 20 million objects. Context is crucial to understanding responsibility and inefficient object connectivity. We summarize memory footprint with help from the dominator relation. Each dominator tree captures unique ownership. Edges between trees capture responsibility. We introduce a set of ownership structures, and quantify their abundance. We aggregate these structures, and use thresholds to identify important aggregates. We introduce the ownership graph to summarize responsibility, and backbone equivalence to aggregate patterns within trees. Our implementation quickly generates concise summaries. In two minutes, it generates a 14-node ownership graph from 29 million objects. Backbone equivalence identifies a handful of patterns that account for 80% of a tree's footprint. © Springer-Verlag Berlin Heidelberg 2006.