special issue: parallel and distributed computing (europar 2005)

2
CONCURRENCY AND COMPUTATION: PRACTICE AND EXPERIENCE Concurrency Computat.: Pract. Exper. 2007; 19:2183–2184 Published online 20 September 2007 inWiley InterScience (www.interscience.wiley.com). DOI: 10.1002/cpe.1267 Special Issue: Parallel and Distributed Computing (EuroPar 2005) Parallel and distributed computing has been changing since the 1980s in many significant dimen- sions. Improvements in computation, storage, and communication infrastructures have enabled new classes of parallel and distributed applications. Evolutions at the computer system and architecture levels have led to the development of a diversity of platforms, ranging from supercomputers, shared- and distributed-memory multiprocessors, local- and wide-area networks, to heterogeneous large- scale distributed grid environments. Such changes have motivated important advances in systems software, tools, and environments, and in the programming models, languages, and methods for parallel and distributed computing. Such diversity of dimensions has also required a change in perspective of the software developer, in order to face the difficulties of meeting the correctness and performance specifications, as required by each specific application. For instance, the techniques and tools for parallel program debugging and performance tuning for a tightly coupled multiprocessor are quite distinct from the ones for debugging and performance evaluation of a large-scale distributed grid application. The changes in paradigms relate to multiple dimensions, such as the degrees of concurrency and non-determinism: applications are intrinsically concurrent and distributed, with data-dependent behavior; the variable, dynamic, and irregular nature of the data sets, and their increasingly large scale; the variable degrees of decentralization of data and control, not only to adapt to the geograph- ical distribution of the application components but also to meet the increasing demands on availability and reliability, as well as to address the mobility concerns of computation, data, and devices; the need to address new concepts (or to rethink and adapt older concepts) for computation, storage, and communication. Owing to the above, in the recent past there has been an increasing integration and/or inter- operation of apparently distinct (but which are in fact complementary in many perspectives) models and technologies from parallel and distributed computing domains: ranging across parallel multi- threaded components, Web services, mobile, and Grid computing models. It is clear that such a change and evolution requires new solutions from the perspective of support tools and environments for parallel and distributed application development, and the correspond- ing methodologies for ensuring program correctness, and for evaluating and predicting program behaviors, namely performance and reliability. Copyright © 2007 John Wiley & Sons, Ltd.

Upload: jose-c-cunha

Post on 11-Jun-2016

212 views

Category:

Documents


0 download

TRANSCRIPT

CONCURRENCY AND COMPUTATION: PRACTICE AND EXPERIENCEConcurrency Computat.: Pract. Exper. 2007; 19:2183–2184Published online 20 September 2007 inWiley InterScience (www.interscience.wiley.com). DOI: 10.1002/cpe.1267

Special Issue:Parallel and DistributedComputing (EuroPar 2005)

Parallel and distributed computing has been changing since the 1980s in many significant dimen-sions. Improvements in computation, storage, and communication infrastructures have enabled newclasses of parallel and distributed applications. Evolutions at the computer system and architecturelevels have led to the development of a diversity of platforms, ranging from supercomputers, shared-and distributed-memory multiprocessors, local- and wide-area networks, to heterogeneous large-scale distributed grid environments. Such changes have motivated important advances in systemssoftware, tools, and environments, and in the programming models, languages, and methods forparallel and distributed computing.Such diversity of dimensions has also required a change in perspective of the software developer,

in order to face the difficulties of meeting the correctness and performance specifications, as requiredby each specific application. For instance, the techniques and tools for parallel program debuggingand performance tuning for a tightly coupled multiprocessor are quite distinct from the ones fordebugging and performance evaluation of a large-scale distributed grid application.The changes in paradigms relate to multiple dimensions, such as

• the degrees of concurrency and non-determinism: applications are intrinsically concurrent anddistributed, with data-dependent behavior;

• the variable, dynamic, and irregular nature of the data sets, and their increasingly large scale;• the variable degrees of decentralization of data and control, not only to adapt to the geograph-ical distribution of the application components but also to meet the increasing demands onavailability and reliability, as well as to address the mobility concerns of computation, data,and devices;

• the need to address new concepts (or to rethink and adapt older concepts) for computation,storage, and communication.

Owing to the above, in the recent past there has been an increasing integration and/or inter-operation of apparently distinct (but which are in fact complementary in many perspectives) modelsand technologies from parallel and distributed computing domains: ranging across parallel multi-threaded components, Web services, mobile, and Grid computing models.It is clear that such a change and evolution requires new solutions from the perspective of support

tools and environments for parallel and distributed application development, and the correspond-ing methodologies for ensuring program correctness, and for evaluating and predicting programbehaviors, namely performance and reliability.

Copyright © 2007 John Wiley & Sons, Ltd.

2184 EDITORIAL

The non-determinism and dynamic nature of the computations, and the potentially large scale ofmodern applications in terms of the input and generated data sets, often requires completely newmethods for analysis and prediction of behavior, for example, as derived from the statistics andinformation theory domains, as based on machine-learning approaches.It is such a new scenario that motivates new advances, such as the ones reported in this special

issue. In fact, the invited papers presented in this special issue originated from improvements,developments, and extensions of previous works that were presented at the EuroPar 2005 conference,held at Universidade Nova de Lisboa, in August 2005. Of course, it would not have been possibleto include, in a single issue, a more extensive list of the topics that are traditionally addressed by theEuroPar conference, and many important topics could not be covered in this issue. Nevertheless,overall, the selected papers provide a representative combination of works that illustrate some ofthe above-mentioned concerns.The first three papers address issues on performance analysis and prediction. The paper by

Strohmaier and Shan discusses benchmarks to capture global data transfer and describes perfor-mance considering temporal and spatial locality concepts which are used to compare distinct archi-tectural designs. The paper by Mohr et al. proposes the evolution of an existing tracing event-basedtool performance analysis of new parallel programming constructs dealing with data decomposition,communication, and synchronization. The third paper, by McKee et al., applies machine learningto support the performance prediction of large-scale data and multi-domain parameter spaces.The paper by Drozdowski and Lawenda presents a study on the use of optimization algorithms

for the multi-installment divisible load processing for heterogeneous distributed systems wherecommunication and computation speeds are used to adjust the sizes of the computation parts.The following papers in the issue all address distributed computing issues.The paper by Junqueira and Marzullo presents a study on replication predicates, which provide

lower bounds on process replication when there are faulty processes, considering the case of de-pendent failures. The paper by Barreto and Ferreira presents an optimistic replication protocol forweakly connected replicas. The paper by Proietti and Guala presents an efficient truthful mecha-nism for computing a common network topology. The paper by Voulgaris et al. exploits a proactiveapproach to facilitate searching by building information on semantic relationships between peers,based on an epidemic protocol to group peers with similar content.The guest editors of this special issue would like to express their in-depth gratitude to all authors,

to all external reviewers, and to Geoffrey Fox, Luc Moreau, Adele Hawksworth, and Tim Williamsfor their efforts in making this issue possible and for their patience.

JOSE C. CUNHA and PEDRO D. MEDEIROS

Departamento de Informatica,Faculdade de Ciencias e Tecnologia,

Universidade Nova de Lisboa,Quinta da Torre, 2829-516 Caparica, Portugal

Copyright q 2007 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. 2007; 19:2183–2184DOI: 10.1002/cpe