Andreas Quick

Distributed performance monitoring: methods, tools, and applications

A method for analyzing the functional behavior and the performance of programs in distributed systems is presented. We use hybrid monitoring, a technique which combines advantages of both software monitoring and hardware monitoring. The paper contains a description of a hardware monitor and a software package (ZM4/SIMPLE) which make our concepts available to programmers, assisting them in debugging and tuning of their code. A short survey of related monitor systems highlights the distinguishing features of our implementation. As an application of our monitoring and evaluation system, the analysis of a parallel ray tracing program running on the SUPRENUM multiprocessor is described. It is shown that monitoring and modeling both rely on a common abstraction of a systems dynamic behavior and therefore can be integrated to one comprehensive methodology. This methodology is supported by a set of tools. >

Automatic scalability analysis of parallel programs based on modeling techniques

When implementing parallel programs for parallel computer systems the performance scalability of these programs should be tested and analyzed on different computer configurations and problem sizes. Since a complete scalability analysis is too time consuming and is limited to only existing systems, extensions of modeling approaches can be considered for analyzing the behavior of parallel programs under different problem and system scenarios. In this paper, a method for automatic scalability analysis using modeling is presented. Initially, we identify the important problems that arise when attempting to apply modeling techniques to scalability analysis. Based on this study, we define the Parallelization Description Language (PDL) that is used to describe parallel execution attributes of a generic program workload. Based on a parallelization description, stochastic models like graph models or Petri net models can be automatically generated from a generic model to analyze performance for scaled parallel systems as well as scaled input data. The complexity of the graph models produced depends significantly on the type of parallel computation described. We present several computation classes where tractable graph models can be generated and then compare the results of these automatically scaled models with their exact solutions using the modeling tool PEPP.

Performance evaluation of a communication system for transputer-networks based on monitored event traces

Most parallel applications (e.g. image processing, mtdtigrid algorithms) in a Transputer-network require a lot of communication between the processing nodes. For such applications the communication system TRACOS was developed to support data transfer between random Transputers in the network. To marimize the performance of the parallel system, its dynamic internal behavior has to be analyzed. For this purpose event-driven monitoring is an appropriate technique. It reduces the dynamic behavior of the system to sequences of events. They are recor&d by a monitor system and stored as event traces. In this paper the communication system TRACOS and its performance evaluation based on monitored event traces are presented. First a synthetic workload was instrumented and monitored with the distributed hardware monitor ZM4. The results showed that the performance of TRACOS is poor for packets smaller than 4 Kbyte. Therefore, TRACOS itself was instrumented and monitored to get insight into the interactions and interdependencies of all TRACOS processes. Based on the monitoring resuhs, TRACOS could be improved which led to a performance increase of 25T0.

Performance Evaluation of Parallel Programs — Modeling and Monitoring with the Tool PEPP

There are many possibilities how to parallelize an algorithm and how to map a program onto a parallel or distributed system. Performance models help to predict which implementation and which mapping are the best for a given algorithm and for a given computer configuration. Stochastic graph modeling is an appropriate method, since the execution order of tasks, their runtime distribution, and branching probabilities are represented. In this paper the modeling capabilities and the analysis techniques implemented in our tool PEPP are presented. In order to obtain relevant modeling results, measured and not only estimated model parameters are needed. They can be obtained through monitoring existing programs. A method to carry out monitoring efficiently is model-driven monitoring: model tasks are mapped onto their corresponding program activities which allows systematic and automatic program instrumentation. Model parameters can easily be calculated since the set of events is the same in modeling and monitoring. A model without timing information can be enhanced to a performance model with realistic parameters.

Synchronisierte Software-Messungen zur Bewertung des dynamischen Verhaltens eines UNIX-Multiprozessor-Betriebssystems

Bei der Entwicklung parallel ablaufender Programme in Muldprozessorsystemen benotigt man Informationen daruber, wie und wo das Programm im System bearbeitet wird. Da aber gerade das Zusammenspiel und die Wechselwirkung aller Systemkomponenten in den meisten Fallen nicht von ausen sichtbar ist, mus dem Entwickler ein Einblick in das interne, dynamische Verhalten des Systems vermittelt werden. Um einen solchen Einblick in das dynamische Verhalten eines Multiprozessorsystems zu bekommen, ist das Beobachten des aktiven Systems (Monitoring) ein geeignetes Hilfsmittel. Im folgenden wird eine Methode vorgestellt, durch Definition relevanter Ereignisse und ihrer zeitgerechten Beobachtung das Verhalten eines Multiprozessor systems nach ausen hin sichtbar zu machen und in Ereignisspuren zu erfassen. Diese sind dann Grundlage fur eine Analyse des Systemverhaltens. Anhand einiger Beispiele wird exemplarisch fur ein UNIX-Multiprozessor-Betriebssystem die Auswertung von Ereignisspuren erlautert.

Stochastic modeling of scaled parallel programs

Testing the performance scalability of parallel programs can be a time consuming task, involving many performance runs for different computer configurations, processor numbers, and problem sizes. Ideally, scalability issues would be addressed during parallel program design, but tools are not presently available that allow program developers to study the impact of algorithmic choices under different problem and system scenarios. Hence, scalability analysis is often reserved to existing (and available) parallel machines as well as implemented algorithms. In this paper we propose techniques for analyzing scaled parallel programs using stochastic modeling approaches. Although allowing more generality and flexibility in analysis, stochastic modeling of large parallel

Integrating Monitoring and Modeling to a Performance Evaluation Methodology

This paper presents a comprehensive methodology for monitoring and modeling parallel and distributed systems systematically. The integration of models, measurements, and evaluators to an efficient set of performance evaluation tools is described. Three typical tools are presented. One of them is the distributed hardware and hybrid monitor ZM4, another is the monitor independent and source related event trace interface POETITDL. Both were developed at the Universitat Erlangen-Nurnberg. As a modeling tool stochastic Petri-nets have been used. These tools have been used for analyzing the performance of multiprocessor and multicomputer systems. Here, they are applied in a case study for performance analysis and improvement of a communication subsystem prototype for B(roadband)-ISDN that was developed by IBM’s European Networking Center. The measurement results give some interesting hints concerning the prototype’s architecture which helped to improve the communication subsystem. Measuring the existing communication subsystem was accompanied by models for predicting the performance of modified ones.

Kausalität in Computersystemen

Um die Leistungsfahigkeit paralleler und verteilter Systeme wirklich nutzen zu konnen, mus man die komplizierten Interaktionen zwischen kooperierenden Prozessen verstehen. Dies erfordert unter anderem die Analyse von Kausalbeziehungen, welche ihrerseits den Rahmen fur die moglichen zeitlichen Reihenfolgen aller Ereignisse in einem parallelen und verteilten System bilden.

Leistungsbewertung mit Modellen

Die Bedeutung von Modellen fur die analytische Leistungsbewertung wurde bereits in der Einleitung kurz motiviert. Das zentrale Anliegen ist es dabei, Leistungsaussagen uber (noch) nicht existierende Systeme zu ermoglichen. Das vorliegende Kapitel legt den Schwerpunkt auf ablauforientierte Modellierungsverfahren, die uberwiegend von einem Leistungsbewertungsteam der Universitat Erlangen entwickelt wurden. Die Erlanger Informatik hat sich schon sehr fruh intensiv mit Multiprozessoren befast, und so erklart sich auch bei der Modellierung das vorrangige Interesse an parallelen Programmen und ihrem Ablaufgeschehen. Im Rahmen von Forschungsaktivitaten zum Thema Leistungsbewertung fiel daher die Wahl auf ablauforientierte Verfahren, die die Ablaufdynamik paralleler Programme derart reprasentieren, das sie sowohl Leistungsvorhersagen fur den kompletten Programmlauf auf einer gegebenen Rechnerkonfiguration als auch die Integration von Messung und Modellierung unterstutzen. Der Integrationsgedanke, in diesem Kapitel lediglich angerissen, wird in Kapitel 6 genauer ausgefuhrt.

Die Integration von Monitoring und Modellierung

Die Ablaufe in parallelen und verteilten Systemen sind im allgemeinen so kompliziert, das es sich dringend empfiehlt, die geeigneten potentiellen Ereignisse vor der Messung formal zu spezifizieren. Dieses Kapitel ist einer neuen, von Quick [Qui93] entwickelten Methode zur systematischen und damit auch effizienten Durchfuhrung des Bewertungsprozesses mit Monitoring gewidmet. Diese Methode beruht auf der systematischen Selektion potentieller Ereignisse. Zur formalen Definition potentieller Ereignisse werden ablauforientierte Modelle verwendet, zur Implementierung von Werkzeugen speziell die in Kapitel 5 behandelten stochastischen Graphmodelle.