CallFlow

Résumé du livre

Dynamic call graphs are often used to understand the execution of a software program. CallingContext Trees (CCTs) are fully context aware call graphs that provide an intuitive space to analyze and understand the efficiency of large-scale, complex software programs. To track down performance bottlenecks which are typically unique to a particular code, a given input deck, machine architecture, etc., software developers or performance analysts need to manually explore a CCT to form and verify hypotheses. However, existing tools are not well aligned with the intuition of expert users, nor do they scale for CCTs from a large number of processors. We present CallFlow which employs a flow-based metaphor for visualizing CCTs of large-scale parallel applications. CallFlow treats execution time as a resource being spent during a call chain, which aligns well with the intuition of domain experts. CallFlow is an interactive visual analysis system that provides users with a high-level overview of CCTs together with a number of semantic refinement operations to progressively drill down through the data. We demonstrate the effectiveness of our system with case studies on massively parallel, production simulation codes.