Abstract
Performance prediction models at the source code level are crucial components in advanced optimizing compilers, programming environments, and tools for performance debugging. Compilers and programming environments use performance models to guide the selection of effective code improvement strategies. Tools for performance debugging may use performance prediction models to explain the performance behavior of a program to the user. Finding the best match between a performance prediction model and a specific source–level optimization task or performance explanation task is a challenging problem. The best performance prediction model for a given task is a model that satisfies the precision requirements while including as few performance factors as possible in order to minimize the cost of the performance predictions. In optimizing compilers, the lack of such a cost–effective performance model may make the application of an optimization prohibitively expensive. In the context of a programming environment, marginal performance factors should be avoided since they will obscure reasoning about the observed performance behavior. This paper discusses a new qualitative performance prediction framework at the program source level that automatically selects a minimal set of performance factors for a target system and performance precision requirement. In the context of this paper, a target system consists of a compiler, an operating system, and a machine architecture. The performance prediction framework identifies significant target system and application program parameters that have to be considered in order to achieve the requested precision. Such parameters may include application factors such as number and type of floating point operations, and machine characteristics such as L1 and L2 caches, TLB, and main memory. The reported performance factors can be used by a compiler writer to build or validate a quantitative performance model, and by a user to better understand the observed program performance. In addition, the failure of the framework to produce a model of the desired quality may be an indication that there exists a significant performance factor not considered within the performance framework. Such information is important to guiding a compiler writer or user in a more efficient search for crucial performance factors. Preliminary experimental results for a small computation kernel and a set of twelve target systems indicate the effectiveness of our framework. The target systems for the experiment consisted of four machine architectures (SuperSPARC I-II and UltraSPARC I-II running Solaris 2.5) and three compiler optimization levels (-none, -O3, -depend -fast). Our prototype framework determines different performance models (1) across different precision requirements for the same target system, and (2) across different target systems for the same precision requirement. In addition, the prototype framework leads us to the discovery of a performance factor within the operating system by failing to produce a performance model of a requested precision.