Analysis Performance

The performance of a parallel program is important. If it were not, it would not make sense to go to the trouble and expense of parallelizing the program in the first place. Thus, parallel programs must first be designed so that good performance is possible, then coded to actually achieve that potential. 

We find it useful to distinguish between two types of performance analysis. Performance prediction describes a modeling process that combines the program s behavior with computer system characteristics to yield a quantitative estimate of how fast the program is expected to execute. Performance 

Once a program has been implemented, measurement-based evaluation is useful to refine its implementation by highlighting sources of inefficiency like load imbalance or high communication costs. Performance visualization, based on recorded program behavior, is a particularly powerful method for detecting unexpected behavior within a parallel program. 

At the present time, performance prediction is primarily a conceptual and mathematical process that is not supported by specific tools. Illustrations of the process can be found in many computer science articles that discuss such algorithms.Fiowever, most suppliers of parallel computers and parallel programming software provide some sort of performance evaluation tools matched to their programming environment. There is wide variation in their basic capabilities and assumptions, as well as in the breadth of features and ease of use, but three main classes of capabilities can be identified. 

For data-parallel environments (SIMD programming model), performance can be represented using standard profiling techniques that associate fraction of overall execution time with particular pieces of code. As noted earlier, this programming model has limited utility for the applications we address, and we do not dwell on performance evaluation. 

To analyze the performance of algorithms (a) and (b), we will first obtain expressions for the parallel efficiencies. For these algorithms, which are fully parallelized and involve no communication, load imbalance is the only factor that may contribute significantly to lowering the parallel efficiency, and to predict the parallel performance, we must be able to estimate this load imbalance. Additionally, to make quantitative predictions for the efficiency, it is necessary to collect some statistics for the computational times required for evaluation of the integrals in a shell quartet. 

We will now use these distributions to express the total execution time, that is, the maximum execution time for any one process, for algorithms (a) and (b). This time can be approximated by a sum of the average execution time and a load imbalance term proportional to the standard deviation as 

We have here expressed the load imbalance as the standard deviation a times a factor k p), which is a function of the number of processes, and we will determine the functional form for k p) in section 7.2.2.I. From Eqs. 7.5 and 

To employ the performance models in Eqs. 7.7 and 7.8 for quantitive predictions of the parallel performance, we need an expression for k p) as well as the means and standard deviations associated with the integral computation. We will discuss below how to obtain these parameters, and we will illustrate both the predicted and the actual, measured performance for the two algorithms. 

To obtain an expression for k p) we will assume that the process execution times for both algorithms (a) and (b) form a normal distribution, which is a reasonable assumption (according to the Central Limit Theorem from probability theory) when there is a large number of tasks per process. Assuming a normal distribution with mean /r and standard deviation a, the probability of a process execution time being below fj. + ka can be computed as 5 + jerf(k/V2), where erf denotes the error function. If there are p processes, the probability that all process execution times are below fi+ka is then given as [ -F jerf k/V2)]P. We need Eqs. 7.5 and 7.6 to be fairly accurate estimates for the maximum execution time, and we must therefore choose k such that 

---

Fig. 2. A simplified material thermal performance analysis for a reentry vehicle thermal protection system where = density x surface recession thickness = total aerodynamic heat/heat of ablation ...

MacDonald, R.J. and C.S. Howat, Data Reconciliation and Parameter Estimation in Plant Performance Analysis, AlChE Journal, 34(1), 1988, 1-8. (Parameter estimation)... 

The goal of plant-performance analysis is to develop an accurate understanding of plant operations. This understanding can be used to ... 

The results of plant-performance analysis ultimately lead to a more efficient, safe, profitable operation. 

Historical Definition Plant-performance analysis has been defined as the reconcihation, rec tification, and interpretation of plant... 

Plant-Performance Triangle This view of plant-performance analysis is depicted in Fig. 30-1 as a plant-performance triangle. Figure 30-2 provides a key to the symbols used. 

Extended Plant-Performance Triangle The historical representation of plant-performance analysis in Fig. 30-1 misses one of the principal a ects identification. Identification establishes troubleshooting hypotheses and measurements that will support the level of confidence required in the resultant model (i.e., which measurements will be most beneficial). Unfortunately, the relative impact of the measurements on the desired end use of the analysis is frequently overlooked. The most important technical step in the analysis procedures is to identify which measurements should be made. This is one of the roles of the plant-performance engineer. Figure 30-3 includes identification in the plant-performance triangle. 

Unit layout as installed is the next step of preparation. This may take some effort if analysts have not been involvea with the unit prior to the plant-performance analysis. The equipment in the plant should correspond to that shown on the PFDs and P IDs. Wmere differences are found, analysts must seek explanations. While a hne-by-line trace is not required, details of the equipment installation and condition must be understood. It is particularly useful to correlate the sample and measurement locations and the bypasses shown on the P IDs to those ac tuaUy piped in the unit. Gas vents and liquid (particularly water-phase) discharges may have been added to the unit based on operating experience out not shown on the P IDs. While these flows may ultimately be small within the context of plant-performance an ysis, they may have sufficient impact to alter conclusions regarding trace component flows, particularly those that have a tendency to build in a process. 

These potential sampling problems must be solved in advance of the unit test. The conclusions drawn from any unit test are strongly affected by the accuracy of the sampling methods and the resultant analyses. Methods should be discussed and practiced before the actual unit test. Analysts should use the trial measurements in prehm-inary plant-performance analysis to ensure that the results will be use-bil during the actual unit test. 

Consider Fig. 30-10. This is a single unit process with one input and two output streams. The goal for plant-performance analysis is to understand accurately the operation of this unit. 

Plant-performance analysis reqmres the proper analysis of limited, uncertain plant measurements to develop a model of plant operations for troubleshooting, design, and control. 

The purpose of the plant-performance analysis is to operate on the set of measurements obtained, subject to the equipment constraints to troubleshoot to develop models or to estimate values for model parameters. 

Measurement Selection The identification of which measurements to make is an often overlooked aspect of plant-performance analysis. The end use of the data interpretation must be understood (i.e., the purpose for which the data, the parameters, or the resultant model will be used). For example, building a mathematical model of the process to explore other regions of operation is an end use. Another is to use the data to troubleshoot an operating problem. The level of data accuracy, the amount of data, and the sophistication of the interpretation depends upon the accuracy with which the result of the analysis needs to oe known. Daily measurements to a great extent and special plant measurements to a lesser extent are rarelv planned with the end use in mind. The result is typically too little data of too low accuracy or an inordinate amount with the resultant misuse in resources. 

This is a formidable analysis problem. The number and impact of uncertainties makes normal pant-performance analysis difficult. Despite their limitations, however, the measurements must be used to understand the internal process. The measurements have hmited quahty, and they are sparse, suboptimal, and biased. The statistical distributions are unknown. Treatment methods may add bias to the conclusions. The result is the potential for many interpretations to describe the measurements equaUv well. 

Whitney, W.J., Comparative Study of Mixed and Isolated Flow Methods for Cooled Turbine Performance Analysis, NASA, TM X-1572, 1968. 

Performance analysis is not only extremely important in determining overall performance of the cycle but in also determining life cycle considerations of various critical hot section components. 

Improvement of equipment effectiveness. This should start with a detailed design review of the plant machinery. A performance analysis of the plant could point to a specific area known to have problems (i.e., section of plant) must be selected and focused on, project teams should be formed and assigned to each train. An analysis should be conducted that address the following ... 

The section on Performance provides a handy nomograph for quick cooling tower evaluation. More detailed analysis requires the use of transfer units. The performance analysis then asks two questions ... 

In quadrupole-based SIMS instruments, mass separation is achieved by passing the secondary ions down a path surrounded by four rods excited with various AC and DC voltages. Different sets of AC and DC conditions are used to direct the flight path of the selected secondary ions into the detector. The primary advantage of this kind of spectrometer is the high speed at which they can switch from peak to peak and their ability to perform analysis of dielectric thin films and bulk insulators. The ability of the quadrupole to switch rapidly between mass peaks enables acquisition of depth profiles with more data points per depth, which improves depth resolution. Additionally, most quadrupole-based SIMS instruments are equipped with enhanced vacuum systems, reducing the detrimental contribution of residual atmospheric species to the mass spectrum. 

Performance analysis procedure Design review procedure... 

Customer complaints Warranty claims Failure analysis reports Process capability studies Service reports Concessions Change requests Subcontractor assessments Performance analysis Deviations and waivers Contract change records Quality cost data External Quality Audit records... 

---