G.E. Lyon, R. Snelick, R. Kacker, Synthetic-perturbation tuning of MIMD programs, J. Supercomput. 8 (1994) 5-27. (A companion multivariate statistical analysis tool, S-Check, automates much of this. See http://www.scheck.nist.gov/ scheck for free copies.)  Home/Search   Document Not in Database   Summary   Related Articles   Check

....statistical experiment paradigm is well suited for this engineering development phase: measure response = model = predict details: The measured response is a readily accessible global characteristic [1] Program runtime, T , is typical. If necessary, T can be measured by hand with a stopwatch [3]. The model above is not structural it will be an approximate series expansion of the response T treated as a function, T ( 8] Many cases need only a few expansion terms to get a serviceable model. 3.1 The test explained Let X = Gamma1 denote the usual version of a code section x suspected ....

LYON G. et al., Synthetic-perturbation Tuning of MIMD Programs, The Journal of Supercomputing, 8, n 1, 1994, p. 5-27. [A companion multivariate statistical analysis tool, S-Check, automates much of this paper for C language users. The Web URL is http://www.scheck.nist.gov/scheck.]

....create unnecessary modeling complexity. They showed that the overall variance of execution time due to non deterministic factors is small for most shared memory programs, and presented the potential advantages of using a deterministic model for parallel program performance prediction. Lyon et al. [38] made another claim against stochastic models, arguing that performance analysis is possible at a macro level, where particular details in the system or in the application are largely ignored. They inserted synthetic perturbations in a program, and measured their effects on global performance. The ....

....is stable, and do not attempt to confirm that assumption. The studies of program stability, in general, either try to induce and detect the occurrence of races, like in [18] or seek to discover the sensitive of a certain metric (e.g. execution time) to specific sections in the code, like in [38]. None of these methods addresses the quantification of the potential instability in program behavior, in terms of variability in execution events for repeated runs of the program. In our work, we both induce variations in behavior and quantify such variations, thus providing a concrete basis for ....

[Article contains additional citation context not shown here]

Gordon Lyon, Robert Snelick, and Raghu Kacker. Synthetic-perturbation tuning of MIMD programs. Technical Report NISTIR 5131, National Institute of Standards and Technology, February 1993.

.... program run [27,54,55,61,69,77,80] In this common approach, much of the work is done manually by a knowledgable expert conducting the tuning study, frequently with the use of visualization tools [27,41,51,73] Recently, research has focused on automating the diagnostic process (steps 1, 2, and 3) [20,28,45,50,70]. Several projects [30,25,44,49] are attempting to redefine performance tuning as an interactive run time activity that may perform steering adjustments or debugging fixes as the application runs (step 4) Completing the spectrum, there is current research focused on complete automation of all ....

....to the programmer) than that of ATExpert because Apprentice is not closely tied to a parallelizing compiler. Both of these are post mortem tools, that is, they generate some performance measurements during the application s execution and perform an analysis after execution has completed. S Check [50,71], a tool developed at NIST, uses a partially automated approach to diagnose performance critical parts of large scale applications. The tool uses artificially introduced delays together with the statistical technique known as Design of Experiments or DEX [4] to focus attention on synchronization ....

G. Lyon, R. Snelick, and R. Kacker. Synthetic-perturbation tuning of MIMD programs. The Journal of Supercomputing, 8:5--28, 1994.

....instance, some computing phases are more sensitive than others. The performance of some program segments or run time parameters may change dramatically at some hardware software thresholds. In addition to conventional statistical methods such as synthetic perturbation screening and curve fitting [10], new methods are also introduced to SCALA. Noticeably, based on factorial analysis, we have proposed a methodology for examining latency hiding techniques of hierarchical memory systems [17] This methodology consists of four levels of evaluation. For a set of codes and a set of machines, we ....

Lyon, G., Snelick, R., and Kacker, R. Synthetic-perturbation tuning of mimd programs. Journal of Supercomputing 8, 1 (1994), 5--8.

....series parallel directed acyclic graph and machine resources as service centers in a queueing network model. On several test cases, they obtained very accurate predictions. However, Adve and Vernon [1] suggested recently that stochastic models may create unnecessary modeling complexity. Lyon et al. [6] made another claim against stochastic models, by proposing performance analysis at a macro level, thus ignoring particular details in the systems or in the applications. They inserted synthetic perturbations in a program, and measured their effect on global performance. The major goal was to find ....

....machine and the architectural features of both machines. If the execution graph does not change significantly across machines, we can confidently use the event order on the first machine as a basis for prediction. One possible way to assess program stability relies on time perturbation analysis [6]. The idea is to perturb the original program and verify the effect of such perturbation. Several instrumented versions of the program are executed, each with a specific set of time delays inserted in the code. By comparing the execution graph obtained in each experiment with the original ....

Lyon, G., Snelick, R., and Kacker, R. Synthetic-perturbation tuning of MIMD programs. Tech. Rep. NISTIR 5131, National Institute of Standards and Technology, February 1993.

....complex [1] Perturbation analysis has also been studied in the context of instrumentation intrusion [5, 9] There, however, perturbation models are developed as a tool to compensate perturbations, and approximate the original program behavior when no instrumentation is used. Lyon et al. [3] developed performance analysis by inserting synthetic perturbations in a program, and measuring the effect on global performance. Thus, particular details in the systems or applications can be ignored, and the analysis is conducted at a macro level. 3 Stability of Parallel Programs Our approach ....

Lyon, G., Snelick, R., and Kacker, R. Synthetic-perturbation tuning of MIMD programs. Tech. Rep. NISTIR 5131, National Institute of Standards and Technology, February 1993.

....to analyze the application behavior, but a semiautomatic analyze [14] 5] can help him in two ways: Finding execution bottlenecks and point out how to modify his program to enhance performances, but also highlight hardware critical parameters for the application. We have, for this purpose, adapted [13] a technique called Synthetic Perturbation Analysis (SPA) based on a factorial design experiments [10] For this purpose the code has to be instrumented inserting user defined events that mark the entry and exit of functions or blocks which are going to be analyzed. The functionality of this ....

G. Lyon, R. Snelick and R. Kacker, "Synthetic-Perturbation Tuning of MIMD Programs", The Journal of Supercomputing, Vol. 8, pp 5-28, 1994.

....response is its patches in a one minute trial. Given that the host system is readily available but perhaps not well understood, empirical scalability testing can be highly attractive. It is especially powerful at screening software factors to find those associated with performance problems [7]i. 3.1. A handy test The following is an easy method for checking code components for scalability. Source code is not needed, only the ability to patch in delays. The actual system must be available for running the code. Table 1 has measurements of a parallel version of SLALOM, where ....

....is not something one can expect on all scalable systems [9] Complex interactions between specimen and host upset both methods. A performance region with a pronounced non linear scaling response might throw EST off track. Closely spaced host sizes will help, but the cost of using EST rises [7]. MK can always incorporate further complexity into its model and simulation runs, but this becomes more expensive. Once constructed, F (a b) a F (c d) c. Setup code scales. The right side of Eq. 3) expresses how performance R changes as X and 5 assume different settings, as in Table 1. ....

G.E. Lyon, R. Snelick, R. Kacker, Synthetic-perturbation tuning of MIMD programs, J. Supercomput. 8 (1994) 5-27. (A companion multivariate statistical analysis tool, S-Check, automates much of this. See http://www.scheck.nist.gov/ scheck for free copies.)

No context found.

G. Lyon, R. Snelick and R. Kacker, `Synthetic-perturbation tuning of MIMD programs', The Journal of Supercomputing, 8, 5--27 (1994). Also in short form as `Time-perturbation tuning of MIMD programs', in R. Pooley and J. Hillston (eds.), Computer Performance Evaluation '92: Modelling Techniques and Tools, Edinburgh University Press Ltd, 1993, pp. 155--167.

No context found.

G. Lyon, R. Snelick and R. Kacker, `Synthetic-perturbation tuning of MIMD programs', The Journal of Supercomputing, 8, 5--27 (1994).

No context found.

G.Lyon, R.Snelick, R.Kacker, "Synthetic-Perturbation Tuning of MIMD Programs," NISTR 5131, February 1993.

....code (a simple quicksort program) is under the example directory in the top level of the distribution. Start S Check while in this directory. As we go through the example, user actions will be outlined in highlighted boxes. It is assumed that the reader is familiar with basic S Check concepts [3,4,5]. Below is a listing of the steps and procedures required by S Check to complete its analysis of your code. Some steps are trivial, requiring just a simple user response or are completely automated by S Check. Others involve moderate input and interaction from the user. select experiment name ....

G. Lyon, R. Snelick, and R. Kacker, Synthetic-perturbation tuning of MIMD programs, Journal of Supercomputing 8(1)(1994) 5-8.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC