USING SYSTEM EXECUTION EXECTION TRACES TO ANALYZE PERFORMANCE PROPERTIES OF SOFTWARE SYSTEMS

Abstract

System execution traces (execution logs) are traditionally used to evalu-ate functional properties of a software system. Prior research, however, has shown the usefulness of system execution traces in evaluating software sys-tem performance properties. Due to the complexity and verboseness of a system execution trace, however, higher-level abstractions, e.g., dataflow models are required to support such evaluation. Our current research effort therefore has focused on extending this dataflow model based system per-formance analysis in two folds. In one aspect, we have considered adapting the dataflow model when the system execution trace does not contain prop-erties required to support performance analysis. In the other aspect, we have developed techniques to auto-generate the supporting dataflow model from a system execution trace. The second aspect is critical because it is hard to manually craft a dataflow model for large and complex software sys-tems, especially distributed software systems. The tool and technique we developed for adapting the dataflow model is called System Execution Trace Adaptation Framework (SETAF); whereas, the tool and technique for auto-generating the dataflow model from a system execution trace is called Dataflow Model Auto Constructor (DMAC). Our cur-rent results from applying SETAF to several open-source production software applications show that there are 3 main patterns for adapting system execu-tion traces to support performance analysis if they do not already have the required properties. Likewise, our current results from applying DMAC to the same software applications show that DMAC is able to auto-construct a data-flow model that covers up to 95% of the originating system execution trace. This work was sponsored in part by the Australian Defense Science and Technology Organi-zation (DSTO)