Efstathios Papaefstathiou

Pace--A Toolset for the Performance Prediction of Parallel and Distributed Systems

This paper describes a methodology that provides detailed predictive performance information throughout the software design and implementation cycles. It is structured around a hierarchy of performance models that describe the computing system in terms of its software, parallelization, and hardware components. The methodology is illustrated with an implementation, the performance analysis and characterization environment (PACE) system, which provides information concerning execution time, scalability, and resource use. A principal aim of the work is to provide a capability for rapid calculation of relevant performance numbers without sacrificing accuracy. The predictive nature of the approach provides both pre and post implementation analyses and allows implementation alternatives to be explored prior to the commitment of an application to a system. Because of the relatively fast analysis times, these techniques can be used at runtime to assist in application steering and scheduling with reference to dynamically changing systems and metacomputing.

Performance modeling of parallel and distributed computing using PACE

There is a wide range of performance models being developed for the performance evaluation of parallel and distributed systems. A performance modelling approach described in this paper is based on a layered framework of the PACE methodology. With an initial implementation system, the model described by a performance specification language, CHIP/sup 3/S, can provide a capability for rapid calculation of relevant performance information without sacrificing accuracy of predictions. An example of the performance evaluation of an ASCI kernel application, Sweep3D, is used to illustrate the approach. The validation results on different parallel and distributed architectures with different problem sizes show a reasonable accuracy (approximately 12% error at most) can be obtained, allows cross-platform comparisons to be easily undertaken, and has a rapid evaluation time (typically less than 2s).

Application Execution Steering using On-the-Fly Performance Prediction

The execution of an application on a high performance system requires parameters concerning the problem in hand, and those that determine the system mapping, to be specified by a user. The system parameters are typically used to minimise the execution time. However, by the coupling of a performance model with an application, system parameters can be determined without user intervention. In the work presented here, a novel performance prediction system has been used to provide suitable performance models which can determine application mapping parameters, code execution decisions, and system choices on-the-fly. An example compact application of a convolution is used to illustrate the approach for automatically choosing the actual code to be executed, and the number of workstations in a cluster to be utilised. The performance prediction system is shown to have a reasonable accuracy (approximately 10% error), with a rapid evaluation time (typically < 2s).

Is predictive tracing too late For HPC users

An underlying goal in the use of high performance systems is to apply the complex resources to achieve rapid application execution times. It is often the case that performance issues are considered late in the application development when major design choices and system choices have already been finalised. Performance tuning tools, including parallel monitoring environments, are useful in these late stages providing a means in which to investigate and visualise the performance effects. However, during the development of an application, certain issues are typically decided upon without reference to their impact on performance (e.g. in the choice of a numerical implementation, or in the choice of a possible mapping to the system). There is a clear need for the study of performance at each and every stage of the development of high performance applications.

A Layered Approach to Parallel Software Performance Prediction: A Case Study

An approach to the characterisation of parallel systems using a structured layered methodology is described here. The aim of this is to produce accurate performance predictions which maybe used to influence the choice of machines and investigate implementation trade-offs. The methodology described enables the separate characterisation of both application, and parallel machine to be developed independently but integrated through an intermediary layer encompassing mapping and parallelisation techniques. The layered approach enables characterisations which are modular, re-usable, and can be evaluated using analytical techniques. The approach is based upon methods introduced in Software Performance Engineering (SPE) and structural model decomposition but due to its modular nature, takes less time for development. A case study in image synthesis is considered in which factors from both the application and parallel system are investigated, including the accuracy of predictions, the parallelisation strategy, and scaling behaviour.

Design of a performance technology infrastructure to support the construction of responsive software

The construction of software that meets its performance objectives is a challenging task considering the distributed nature of modern architectures. Although substantial progress is being made in areas of performance technology, a breakthrough in the wider use of prediction tools during the software development process still remains to be seen. Lack of integration and standardization is one of the reasons for the slow adoption of the performance technology. This paper presents the design of a Performance Technology Infrastructure (PTI), an integration environment for hardware models, workload descriptions, and performance analysis tools. PTI includes a workload specification library, a model evaluation engine, and an interface to external hardware models. PTI components interact with XML scripts based on the syntax of predefined schemas that can be extended to include the requirements of new components incorporated into the system. A prototype implementation of a performance prediction tool is also introduced. It includes third party hardware models, a performance specification language, and an analysis tool. Performance predictions for the Sweep3D application running on a PC cluster are obtained and analyzed to demonstrate the tool’s capabilities.

Optimisation of application execution on dynamic systems

In this paper, we demonstrate the impact of using a dynamic (on-the-fly) performance prediction tool-set, PACE, for optimising application execution on dynamic systems. The need for steering the application execution arises from the ever-growing use of distributed and GRID systems. The unquestionable aim to overcome bottleneck problems, allocation, and performance degradation due to shared CPU time has prompted many investigations into the best way in which the performance of an application can be enhanced. In this work, we present a novel approach to dynamically optimise the performance of an application. An example application, the FFTW (the fastest Fourier transform in the west), is used to illustrate the approach which itself is a novel method that optimises the execution of an FFT. It is shown that performance prediction can provide the same quality of information as a measurement process for application optimisation but in a fraction of the time and thus improving the overall application performance.

PASE: A performance analysis simulation environment

Abstract Performance simulation is one of the approaches used to estimate the quantitative behaviour of inter-connection networks of parallel and distributed systems. This paper presents both a discrete-event network simulation environment for multiprocessor systems, and a methodology for simulation. The simulator provides A Network Initiated Simulation Oriented Language (ANISOL) which allows the user to define environmental factors, network topologies and the load sharing strategies in both homogeneous and heterogeneous systems. The simulator is flexible enough to simulate various interconnection network topologies and processing environments. Features required to simulate massively parallel processors, such as fault tolerance, are also included. This simulation environment has been used to study a range of different architectures and preliminary results have shown that the accuracy of the outputs is comparable to other techniques.

Use of Performance Technology for the Management of Distributed Systems

This paper describes a toolset, PACE, that provides detailed predictive performance information throughout the implementation and execution stages of an application. It is structured around a hierarchy of performance models that describes distributed computing systems in terms of its software, parallelisation and hardware components, providing performance information concerning expected execution time, scalability and resource use of applications. A principal aim of the work is to provide a capability for rapid calculation of relevant performance numbers without sacrificing accuracy. The predictive nature of the approach provides both pre- and post- implementation analyses, and allows implementation alternatives to be explored prior to the commitment of an application to a system. Because of the relatively fast analysis times, these techniques can be used at run-time to assist in application steering and efficient management of the available system resources.

Run-Time Optimization Using Dynamic Performance Prediction

With the rapid expansion in the use of distributed systems the need for optimisation and the steering of application execution has become more important. The unquestionable aim to overcome bottle-neck problems, allocation, and performance degradation due to shared CPU time has prompted many investigations into the best way in which the performance of an application can be enhanced. In this work, we demonstrate the impact of using a Performance Prediction Toolset, PACE, which can be used in Dynamic (On-The-Fly) decision making for optimising application execution. An example application, the FFTW (The Fastest Fourier Transform in the West), is used to illustrate the approach which itself is a novel method that optimises the execution of an FFT. It is shown that performance prediction can provide the same quality of information as a measurement process for application optimisation but in a fraction of the time and thus improving the overall application performance.