英语翻译1 IntroductionWith the advent of multi-core processors a

英语翻译
1 Introduction
With the advent of multi-core processors and the growing popularity of local
cluster installations,better understanding of parallel applications behaviour becomes a necessity.The plans to build a sustainable HPC ecosystem for Europe will make hardware installations needed for execution of large-scale parallel applications even more accessible.It can be argued that the raising popularity of parallelization results in the dare need of methods and tools capable of automatic analysis and prediction of parallel applications efficiency.Be it a small cluster or a pan-European grid installation,the possibility to determine efficiency of parallel applications will always be welcomed,if not required.Performance evaluation can be employed to raise the usefulness of all the parallel and distributed processing environments.The results of the evaluation are suitable for resource allocation,load balancing,on-demand reservation of dynamic resources,quality of service negotiations,runtime and wait time predictions,etc.During parallel program performance evaluation various metrics are used - the runtime,speedup or efficiency are all usually measured using the wall-clock time [6].Although the wall-clock time is indisputably one of the most important factors to be measured,it can be shown that concentrating only on this single value can hide important information.Basing on the decomposition of the execution time between the time devoted to the computations and the overhead time,the Execution Time Decomposition (ETD) [1] based analysis can be carried out,providing not only insight into the application execution but also possibility to perform fast and accurate performance estimations.The accuracy of techniques based on internal and external measurements is evaluated and compared to the results achieved with the classical analysis.The structure of the paper is as follows.Section 2 briefly describes different techniques and metrics used during performance evaluation.The next section is dedicated to the description of the ETD approaches analyzed in the paper.The fourth section discusses the experiments performed and the achieved results.The final section concludes the paper and presents the future work plans.
.The performance achieved by a parallel application depends not only on the application features but it also depends on the interactions between the hardware and software resources of the system where the application is being executed.Such application characteristics as algorithmic structure,input parameters,problem size,influence these interactions by determining how the application exploits the available resources and the allocated processors.It means that when designing the parallel application the goal of the design process is not to optimize a single metrics like for example speed,that is enough in case of sequential applications.The good design has to take into consideration a problem specific function of execution time,memory requirements,implementation cost,and others.In general the performance analysis
can be carried out analytically or through experiments。In the first case we call it performance modeling,when in the second performance measurement.