Requirements for Automatic Performance Analysis - APART Technical Report

This report discusses the requirements for automatic performance analysis tools. The discussion proceeds by first examining the nature and purpose of performance analysis. This results in an identification of the sources of performance data available to the analysis process and some properties of the process itself. Consideration is then given to the automation of the process. Many environmental factors affecting the performance analysis process are identified leading to the definition of a scenario space in which users operate. The requirements associated with some scenarios defined by this space are then explored.Performance is characterised in terms of a number of performance properties associated with low level, hardwareoriented information and also with high level diagnostic statements. Performance properties are defined in terms of the performance data. Performance analysis requires the detection of relevant properties in the execution of an application in the regions of the source code in which they occur along with a statement of their impact on performance. The requirements for defining the data model, detecting properties and locating and quantifying their effect on performance are summarised in a set of high level requirements and requirements derived from these.The process of automatic performance analysis is summarised as being one of query formulation (describing conditions for the existence of properties etc.) and query optimisation (using suitable strategies to search for properties in performance data which may be gathered on demand).An object-oriented data model is suggested as a suitable vehicle for representing all possible performance data, and the requirement for formal languages to describe performance properties, and capture the associated dynamic search process for them, is stated