Author(s): Assunção, Luís Manuel da Costa
Date: 2016
Persistent ID: http://hdl.handle.net/10362/19975
Origin: Repositório Institucional da UNL
Subject(s): Scientific workflows; Parallel and distributed computing; Dynamic reconfiguration; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática
In the last decade we witnessed an immense evolution of the computing infrastructures in terms of processing, storage and communication. On one hand, developments in hardware architectures have made it possible to run multiple virtual machines on a single physical machine. On the other hand, the increase of the available network communication bandwidth has enabled the widespread use of distributed computing infrastructures, for example based on clusters, grids and clouds. The above factors enabled different scientific communities to aim for the development and implementation of complex scientific applications possibly involving large amounts of data. However, due to their structural complexity, these applications require decomposition models to allow multiple tasks running in parallel and distributed environments. The scientific workflow concept arises naturally as a way to model applications composed of multiple activities. In fact, in the past decades many initiatives have been undertaken to model application development using the workflow paradigm, both in the business and in scientific domains. However, despite such intensive efforts, current scientific workflow systems and tools still have limitations, which pose difficulties to the development of emerging large-scale, distributed and dynamic applications. This dissertation proposes the AWARD model for scientific workflows with parallel and distributed computing. AWARD is an acronym for Autonomic Workflow Activities Reconfigurable and Dynamic. The AWARD model has the following main characteristics. It is based on a decentralized execution control model where multiple autonomic workflow activities interact by exchanging tokens through input and output ports. The activities can be executed separately in diverse computing environments, such as in a single computer or on multiple virtual machines running on distributed infrastructures, such as clusters and clouds. It provides basic workflow patterns for parallel and distributed application decomposition and other useful patterns supporting feedback loops and load balancing. The model is suitable to express applications based on a finite or infinite number of iterations, thus allowing to model long-running workflows, which are typical in scientific experimention. A distintive contribution of the AWARD model is the support for dynamic reconfiguration of long-running workflows. A dynamic reconfiguration allows to modify the structure of the workflow, for example, to introduce new activities, modify the connections between activity input and output ports. The activity behavior can also be modified, for example, by dynamically replacing the activity algorithm. In addition to the proposal of a new workflow model, this dissertation presents the implementation of a fully functional software architecture that supports the AWARD model. The implemented prototype was used to validate and refine the model across multiple workflow scenarios whose usefulness has been demonstrated in practice clearly, through experimental results, demonstrating the advantages of the major characteristics and contributions of the AWARD model. The implemented prototype was also used to develop application cases, such as a workflow to support the implementation of the MapReduce model and a workflow to support a text mining application developed by an external user. The extensive experimental work confirmed the adequacy of the AWARD model and its implementation for developing applications that exploit parallelism and distribution using the scientific workflows paradigm.
