Workflow management in problem solving environments (PSEs) is an emerging topic that aims to combine both data-oriented and execution-oriented views of scientific experiments, and closely integrate the processes underlying the practice of computational science with the software artifacts constituted by the PSE. This thesis presents a workflow management solution called BREW (BetteR Experiments through Workflow management) that provides functionality along four dimensions: components and installation management, experiment execution management, data management, and (full fledged) workflow management. BREW builds upon EMDAG, a first generation experiment management system designed at Virginia Tech which provided rudimentary facilities for supporting (only) the first two functionalities. BREW provides a complete dynamic workflow management solution wherein the PSE user can compose arbitrary scientific experiments and specify intended dynamic behavior of these experiments to an extent not previously possible. Along with the design details of the BREW system, this thesis identifies important tradeoffs underlying workflow management for PSEs, and presents two case studies involving large-scale data assimilation in bioinformatics experiments. / Master of Science

A problem solving environment (PSE) provides all computational facilities necessary for solving a target class of problems efficiently. PSEs are used primarily for domain-specific problem-solving in science and engineering and aim to ease the burden of advanced scientific computing. Scientific problem solving, however, often involves the use of legacy resources which are difficult to modify or port, and may be distributed on different machines. Existing PSEs provide little support for solving such problems in a generic framework. This thesis investigates the design of a platform-independent system that enables problem solving using legacy resources without having to modify legacy code. It presents Symphony, an open and extensible Java-based framework for composition and manipulation of distributed legacy resources. Symphony allows users to compose visually a collection of programs and data by specifying data-flow relationships among them and provides a client/server framework for transparently executing the composed application. Additionally, the framework is web-aware and helps integrate web-based resources with legacy resources. It also enables programmers to provide a graphical interface to legacy applications and to write visualization components. Symphony uses Sun Microsystems' JavaBeans component architecture for providing components that represent legacy resources. These components can be customized and composed in any standard JavaBeans builder tool. Executable components communicate with a server, implemented using Java Remote Method Invocation mechanism, for executing remote legacy applications. Symphony enables extensibility by providing abstract components which can be extended by implementing simple interfaces. Beans implemented from the abstract beans can act as data producers, consumers or filters. / Master of Science

Nonlinear algebraic equations (NAEs) occur in many areas of science and engineering. The process of solving these NAEs is generally difficult, from finding a good initial guess that leads to a desired solution to deciding on convergence criteria for the approximate solution. In practice, Newton's method is the only robust general-purpose method for solving a system of NAEs. Many variants of Newton's method exist. However, it is generally impossible to know a priori which variant of Newton's method will be effective for a given problem.<p>Many high-quality software libraries are available for the numerical solution of NAEs. However, the user usually has little control over many aspects of what the library does. For example, the user may not be able to easily switch between direct and indirect methods for the linear algebra. This thesis describes a problem-solving environment (PSE) called pythNon for studying the effects (e.g., performance) of different strategies for solving systems of NAEs. It provides the researcher, teacher, or student with a flexible environment for rapid prototyping and numerical experiments. In pythNon, users can directly influence the solution process on many levels, e.g., investigation of the effects of termination criteria and/or globalization strategies. In particular, to show the power, flexibility, and ease of use of the pythNon PSE, this thesis also describes the development of a novel forcing-term strategy for approximating the Newton direction efficiently in the pythNon PSE.

Nonlinear algebraic equations (NAEs) occur in many areas of science and engineering. The process of solving these NAEs is generally difficult, from finding a good initial guess that leads to a desired solution to deciding on convergence criteria for the approximate solution. In practice, Newton's method is the only robust general-purpose method for solving a system of NAEs. Many variants of Newton's method exist. However, it is generally impossible to know a priori which variant of Newton's method will be effective for a given problem.<p>Many high-quality software libraries are available for the numerical solution of NAEs. However, the user usually has little control over many aspects of what the library does. For example, the user may not be able to easily switch between direct and indirect methods for the linear algebra. This thesis describes a problem-solving environment (PSE) called pythNon for studying the effects (e.g., performance) of different strategies for solving systems of NAEs. It provides the researcher, teacher, or student with a flexible environment for rapid prototyping and numerical experiments. In pythNon, users can directly influence the solution process on many levels, e.g., investigation of the effects of termination criteria and/or globalization strategies. In particular, to show the power, flexibility, and ease of use of the pythNon PSE, this thesis also describes the development of a novel forcing-term strategy for approximating the Newton direction efficiently in the pythNon PSE.

An initial-value problem consists of an ordinary differential equation subject to an initial condition. The right-hand side of the differential equation can be interpreted as additively split when it is comprised of the sum of two or more contributing factors. For instance, the right-hand sides of initial-value problems derived from advection-diffusion-reaction equations are comprised of the sum of terms emanating from three distinct physical processes: advection, diffusion, and reaction. In some cases, solutions to initial-value problems can be calculated analytically, but when an analytic solution is unknown or nonexistent, methods of numerical integration are used to calculate solutions. The runtime performance of numerical methods is problem dependent; therefore, one must choose an appropriate numerical method to achieve favourable performance, according to characteristics of the problem. Additive methods of numerical integration apply distinct methods to the distinct contributing factors of an additively split problem. Treating the contributing factors with methods that are known to perform well on them individually has the potential to yield an additive method that outperforms single methods applied to the entire (unsplit) problem. Splittings of the right-hand side can be physics-based, i.e., based on physical characteristics of the problem, such as advection, diffusion, or reaction terms. Splittings can also be based on linearization, called Jacobian splitting in this thesis, where the linearized part of the problem is treated with one method and the rest of the problem is treated with another. A comparison of these splitting techniques is performed by applying a set of additive methods to a test suite of problems. Many common non-additive methods are also included to serve as a performance baseline. To perform this numerical study, a problem-solving environment was developed to evaluate permutations of problems, methods, and their associated parameters. The test suite is comprised of several distinct advection-diffusion-reaction equations that have been chosen to represent a wide range of common problem characteristics. When solving split problems in the test suite, it is found that additive Runge–Kutta methods of orders three, four, and five using Jacobian splitting generally outperform those same methods using physics-based splitting. These results provide evidence that Jacobian splitting is an effective approach when solving such initial-value problems in practice.

Multidisciplinary computational simulations involve interactions between distributed applications, datasets, products, resources, and users. Because the very nature of the simulation software emphasizes a single-computer, small-usership and audience, the kinds of applications that have been developed often are unfriendly to incorporation into a distributed model. However, advances in networking infrastructure, and the natural tendency for information to be geographically distributed place strong requirements on integration of single-computer codes with distributed information sources, as well as multiple computer codes that are geographically distributed in their execution. The hypothesis of this dissertation is that it is possible, via novel integration of Internet, Distributed Computing, and Grid technologies, to create a distributed computational simulation systems that satisfies the requirements of modern multidisciplinary computational simulation systems without compromising functionality, performance, or security of existing applications. Furthermore, such a system would integrate disparate applications, resources, and users and would improve the productivity of users by providing new functionality not currently available. The hypothesis is proved constructively by first prototyping the Enterprise Computational Services framework based on a multi-tier architecture using the Java 2 Enterprise Edition platform and Web Services and then two distributed systems, the Distributed Marine Environment Forecast System and Distributed Simulation System for Seismic Performance of Urban Regions, are prototyped using this enabling framework. Several interfaces to the framework are prototyped to illustrate that the same framework can be used to develop multiple front-end clients required to support different types of users within a given computational domain. The two domain specific distributed environments prototyped using the framework illustrate that the framework provides a reusable common infrastructure irrespective of the computational domain. The effectiveness and utility of the distributed system and the framework are demonstrated by using a representative collection of computational simulations. Additional benefits provided by the distributed systems in terms of new functionality provided are evaluated to determine the impact on user productivity. The key contribution of this dissertation is a reusable infrastructure that could evolve to meet the requirements of next-generation hardware and software architectures while supporting interaction between a diverse set of users and distributed computational resources and multidisciplinary applications.

This thesis investigates several issues in data and computation modeling for scientific problem solving environments (PSEs). A PSE is viewed as a software system that provides (i) a library of simulation components, (ii) experiment management, (iii) reasoning about simulations and data, and (iv) problem solving abstractions. Three specific ideas, in functionalities (ii)-(iv), form the contributions of this thesis. These include the EMDAG system for experiment management, the BSML markup language for data interchange, and the use of data mining for conducting non-trivial parameter studies. This work emphasizes data modeling and management, two important aspects that have been largely neglected in modern PSE research. All studies are performed in the context of S4W, a sophisticated PSE for wireless system design. / Master of Science

Τα τελευταία χρόνια έχουν γίνει σημαντικές προσπάθειες o Πληροφορικός-Επιστήμονας των Υπολογισμών να εκθέσει με εύληπτο τρόπο τη γνώση και εμπειρία του στις κοινότητες εκείνων που θέλουν να κάνουν υπολογισμούς. Κάτι τέτοιο έχει καταστεί δυνατό με την κατασκευή σύνθετων στη δομή, αλλά εύκολων στη χρήση, εργαλείων-περιβαλλόντων υπολογισμού στα οποία κανείς μπορεί με εντελώς φυσικό τρόπο να προδιαγράψει το πρόβλημά του και -ανάλογα με την εμπειρία του- να επέμβει στη ροή επίλυσής του. Τα Περιβάλλοντα Επίλυσης Προβλημάτων (ΠΕΠ) προβάλλουν λοιπόν ως μια πολύ ελκυστική λύση για τον επιστήμονα των εφαρμογών που αναζητεί μια εύχρηστη, ισχυρή και αξιόπιστη πλατφόρμα λογισμικού για τους υπολογισμούς του. Σε πολλές περιπτώσεις αυτοί οι υπολογισμοί είναι πολύ μεγάλης κλίμακας και απαιτούν πολυάριθμους και αποδοτικούς πόρους. Η τιθάσευσή τους σε κάποια έκταση έγινε δυνατή με τη στροφή σε παράλληλες-κατανεμημένες αρχιτεκτονικές, πρόσφατα μεγάλης κλίμακας, με έμφαση στην ευχρηστία, στην ασφάλεια πρόσβασης και στη συνεργατικότητα (Πλέγμα (Grid)). Σε άλλες περιπτώσεις οι πολυπύρηνοι επεξεργαστές που εξοπλίζουν πλέον τους τυπικούς οικιακούς υπολογιστές μας και οι προβλέψεις για αθρόα κλιμάκωση του αριθμού των προσφερόμενων πυρήνων, προτρέπουν σε επαναδιαπραγμάτευση κλασικών αλγορίθμων με στόχευση στην εξαγωγή παραλληλίας, αφού πλέον αυτή μπορεί να απεικονιστεί άμεσα στο διαθέσιμο υλικό. Επιπρόσθετα μια τέτοια στροφή ώθησε και τη διερεύνηση εναλλακτικών μοντέλων υπολογισμού: Το ασύγχρονο μοντέλο υπολογισμού προσφέροντας τη δυνατότητα για εξάλειψη των χρονοβόρων φάσεων συγχρονισμού των πολλαπλών μονάδων επεξεργασίας προβάλλει ως μια ενδιαφέρουσα επιλογή. Συστηματοποιούμε τη μελέτη των Περιβαλλόντων Επίλυσης Προβλημάτων (ΠΕΠ) εντοπίζοντας τους άξονες που χαρακτηρίζουν αυτήν την κατηγορία συστημάτων λογισμικού και υλοποιώντας το Jylab, ένα πρωτότυπο ΠΕΠ με έμφαση στη φορητότητα, την επαναχρησιμοποίηση ελεύθερα διαθέσιμου κώδικα και τη δυνατότητα για ακολουθιακό, παράλληλο και κατανεμημένο υπολογισμό σε πολλαπλές πλατφόρμες. Ειδικότερα, το Jylab περιλαμβάνει υποστήριξη για ασύγχρονο κατανεμημένο υπολογισμό, ανάλυση ιστογραφημάτων και εκτέλεση υπολογισμών στο Πλέγμα (Grid). Αμέσως μετά εισάγουμε το ασύγχρονο μοντέλο υπολογισμού εστιάζοντας σε καίρια ζητήματα όπως η ανάλυση της σύγκλισης, η ανίχνευση του τερματισμού και η υλοποίησή του. Προτείνουμε πιθανοτικό πλαίσιο εντοπισμού της σύγκλισης και διερευνούμε την πολυπλοκότητα του μοντέλου. Στη συνέχεια μελετούμε αλγορίθμους διάταξης των κόμβων ενός γραφήματος, επικεντρώνοντας στον υπολογισμό του διανύσματος του PageRank το οποίο χρησιμοποιεί η Google για να διατάξει τα αποτελέσματα μιας ερώτησης που υποβάλλουμε στη μηχανή αναζήτησής της. Αποδεικνύουμε πως και άλλες μέθοδοι διάταξης, οι οποίες εκφράζονται πρωταρχικά ως δυναμοσειρές ενός τροποποιημένου μητρώου συνδέσμων μπορούν να γραφτούν ως γινόμενα των επαναληπτικών μητρώων που χρησιμοποιούνται στον υπολογισμό του διανύσματος PageRank, αλλά με διαφορετική παράμετρο σε κάθε όρο τους (μέθοδος της πολυπαραμετρικής απόσβεσης). Στη συνέχεια εκθέτουμε την πειραματική συμπεριφορά του ασύγχρονου μοντέλου, όπως αυτή προκύπτει από υλοποιήσεις κυρίως του αλγορίθμου του PageRank, σε διάφορες πλατφόρμες (τοπικά, στη συστάδα υπολογισμών και στο Πλέγμα (Grid)) και με μονάδες εκτέλεσης νήματα ή διεργασίες. To Jylab χρησιμοποιήθηκε εντατικά σε αυτές τις διερευνήσεις και αποδείχτηκε πως όλοι οι πειραματισμοί μπορούν να τεθούν κάτω από ενιαίο πλαίσιο λογισμικού. Επίσης εισάγουμε μια κλάση αλγορίθμων κατανεμημένου υπολογισμού στατιστικών μεγεθών, τους gossip αλγορίθμους, σε κάθε στοιχειώδες βήμα των οποίων μόνο δύο οντότητες επικοινωνούν και υπολογίζουν. Επεκτείνουμε αυτούς τους αλγορίθμους επιτρέποντας σε k > 2 οντότητες να αλληλεπιδρούν ανά βήμα, προσομοιώνουμε τη συμπεριφορά τους και προτείνουμε πρωτόκολλα υλοποίησής τους. / In recent years computational scientists strive to expose their knowledge and experience to the communities of people interested in performing computations. This endeavor focuses on the construction of complex in structure, however simple in use, toolchains and environments in which a researcher can specify his or her problem and - depending on his experience - change its exact solution flow. In many cases these computations necessitate large-scale and performant resources. Harnessing them, to some extent, became possible by turning to parallel-distributed architectures, recently of large scale, emphasizing usability, security in accessing them and collaboration perspectives (Grid). In other cases, the multicore processors, nowadays powering even typical personal computers, coupled with predictions for dramatic increase in the number of available cores in the near future, suggest a reconsideration of classic algorithms aiming at extracting parallelism, since this can be directly mapped to underlying hardware. Additionally, such a move, also fuels the investigation of alternative computation models: The asynchronous computation model, offering the flexibility for the complete removal of time-consuming synchronization phases, is a very interesting option. We study Problem Solving Environments (PSEs) in a systematic manner, specifying the axes characterizing this category of systems of software also implementing Jylab, a prototype PSE emphasizing portability and the reuse of freely available code and enabling sequential, parallel and distributed computing over multiple platforms. More specifically, Jylab includes support for asynchronous distributed computations, Web graph analysis and Grid computing. Then we introduce the asynchronous computation model, focusing in three core subjects, namely its convergence analysis, the termination detection problem and its implementation. We propose a probabilistic framework for convergence detection and explore the complexity of the model. Afterwards, we survey algorithms for ranking the nodes of a graph, focusing on computing the PageRank vector, which is used by Google for ranking the results of a query submitted to its search engine. We prove that a whole class of ranking methods, primarily expressed as a power series of a modified link matrix can be written as products of iterative matrices similar to those used in computing the PageRank vector, albeit with a different damping parameter for each of its terms (multidamping). Next, we present the experimental behavior of the asynchronous model, mainly as applied in computing the PageRank vector, over different platforms (locally, in a computer cluster and over the Grid) using either threads or processes as its units of execution. Jylab was intensively used in these investigations and it was proved that all experimentations can be cast under a unifying software framework. We also introduce a class of algorithms for the distributed computation of statistical quantities, namely gossip algorithms, for which only two entities communicate and compute at each elementary step. We extend these algorithms be permitting k > 2 entities to interact on a per elementary step basis, simulate their behavior and propose protocols for implementing them.