Loading...

Performance Evaluation. Different approaches for evaluating computational performance

Scientific Essay 2013 13 Pages

Computer Science - Commercial Information Technology

Excerpt

Abstract

In this fast moving world, almost every task has become result oriented. Before- hand, planning has to be done in order to assure quick and accurate results. Now a days, super computers are replacing humans nearly in every field. The most common example are robots. As it is rightly said that "If you fail to plan, you plan to fail". Hence it is of utmost importance to check the performance and plan the execution of a system accordingly. To support the above argument this paper throws light on different approaches for evaluating computational performance. Their comparison is also discussed along with some benchmarks which enable users to judge how perfor- mance can be measured. One can be sure of getting good results if early analysis of the system’s operation is done.

1 Introduction

It is vital to understand the behaviour of a system as it’s performance varies in every part of the execution. Today’s super-computers have become surprisingly fast as they finish the assigned tasks within a blink of eyes. These assigned tasks are very much complex. Just imagine the speed of these processors, which execute millions of instructions within fraction of a second. Performance evaluation is the key factor behind this breath-taking speed. Creating such systems involves rigorous and thorough realization of the applica- tions that will run on these systems, which further gives rise to the possible architectures that can meet stringent requirements of the market[23]. This paper discusses various ap- proaches for performance evaluation.

The idea of classification of the various approaches for performance evaluation is taken from the article[23]. There are two approaches namely:Performance Measure-mentandPerformance Modeling. Performance Measurement comes into action when the system of interest is available whereas Performance Modeling is done when the system of interest is not available. In the latter method real system is not accessible hence we

Nishchal Narula

have to build a simulation model. This activity is costly and takes a lot of time. Thus to ensure the quality and durability of the system, performance evaluation is necessary at each stage of the design process. Simulation modelingandAnalytical modelingare the two sub-categories of Performance Modeling whereas Performance Measurement can be further classified intoOn-chip hardware monitoring,Off-chip hardware monitoring,Software monitoringandMicrocoded instrumentaion[23].

ABenchmarkgives you a better insight of how one can evaluate the performance of a given system. This can be dependent on various metrics like throughput and execu- tion time. Although there are many benchmarks/standards available in this computing world however only a few and important of them are discussed in this paper. To ensure the correctness of results during performance measurement and performance modelling, benchmarks which are available for specific applications should be used. Every bench- mark should possess certain features, which are indispensable for the accuracy of results [23]. They are as follows:

i) It should not aggravate system’s performance.
ii) It should be cheap and easily accessible to users. iii) It should be easily modifiable.
iv) It should be easy to use.
v) It should be able to support growing-market’s applications.

The rest of this paper is organized as follows. Section 2 describes the variousap- proaches to performance evaluation. Section 3 discusses someexquisite standards/benchmarksfor various applications. Section 4 presents an example for the better understanding of a working environment -power measurement with the help of simulation. Section 5 high- lights someapplicationsthat where this concept can be further used to enlighten the field of electronics and finally in the Section 6,concluding remarks are given.

2 Approaches to Performance Evaluation

Performance evaluation is regarded as one of the critical activities in the area of software development. One has to dig into every minute detail while addressing this issue. Hence it is indispensable to follow this approach in a planned manner. The methods of performance evaluation are discussed as follows:

2.1 Performance Measurement

Performance measurement is used when the actual system of interest is available and one wants to understand the behaviour of this system. It gives a quantitative measure

illustration not visible in this excerpt

Figure 1: Processor Performance since the 1980’s.

of how well a system works and it determines the usefulness of a system. It gives low flexibility and high accuracy[30]. Generally, measuring performance means calculating the execution time which is the product of Instruction count (Ic), Cycles per instruction (CP I) and Clock period (T)[27].

illustration not visible in this excerpt

(Processor Speed), (Input/Output Speed), (Concurrency of Execution), (Type of Application), (Total Workload of a System), (Choice of Algorithm and Compiler), (Instruction Set Architecture) and (Semiconductor Technology) are some factors that impact the performance of a system[30][27].

Figure 1 shows the processor performance since the 1980’s. One can see that there has been a considerable improvement in the performance of processors. The figure has been taken from[30].

Performance Measurement can be executed by the following processes:

1) Microprocessor On-chip Performance Monitoring Counters Today, majority of the processors integrate on-chip performance monitoring counters to study the behaviour of systems. For instance, in the Intel Pentium processor there are two performance monitoring counters which are read with some instructions such as (RDP M C) - [Read Performance Monitor Counter]. These counters can measure many performance events[23][20].

illustration not visible in this excerpt

Table 1: Some events that can be measured using performance monitoring counters on the Intel Pentium III processor

Table 1 shows some events that can be measured on the Pentium III processor. The contents of this table has been taken from[23].

Some tools are also there to measure performance using the performance monitoring counters. V tuneis the software from Intel which can be used to check the performance using the Intel processor performance counters[13]. The performance of Windows NT can be monitored by theP6P erfplug-in[3]. The performance monitorperfmonis a small hack that uses the on-chip counters on UltraSPARC-I/II processors to gather statistics[5]. Juan Rubio of the Laboratory for Computer Architecture at the University of Texas wrote a counter-reading softwareP M ON[6]to provide a way to read the counters with minimum overhead.

2) Off-chip Hardware Monitoring

AMD has developed a hardware tracing platform for x86 Pentium 4 processors. It interprets and then stores the state of processor which has almost all the vital information about the instructions executing on the processor. However, one has to be careful as this activity can degrade the speed of processor[23]. Merten[25]and Bhargava[31]has also used this technique to measure the performance. Poursepanj and Christie[16]used a Tektronix TLA 700 logic analyzer, which generally has limits on sizes, to examine the 3D graphics workloads on AMD-K6-2 based systems.

3) Software Monitoring

It is an outdated technique and was used when On-Chip performance monitoring coun- ters were not invented. In this technique, the VAX processor throws an exception after every instruction, which causes the processor overhead. Due to this overhead, perfor- mance declines and hence it is not used these days. Ease of use is the only plus-factor of this process[23].

[...]

Details

Pages
13
Year
2013
ISBN (eBook)
9783656414056
ISBN (Book)
9783656414674
File size
550 KB
Language
English
Catalog Number
v212998
Institution / College
University of Paderborn
Grade
Tags
performance evaluation different

Author

Share

Previous

Title: Performance Evaluation. Different approaches for evaluating computational performance