Distributed Management of Grid-based Scientific Workflows

Abstract

Grids and service-oriented technologies are emerging as dominant approaches for distributed systems. With the evolution of these technologies, scientific workflows have been introduced as a tool for scientists to assemble highly specialized applications, and to exchange large heterogeneous datasets in order to automate and accelerate the accomplishment of complex scientific tasks. Several Scientific Workflow Management Systems (SWfMS) have already been designed to support the specification, execution, and monitoring of scientific workflows. Meanwhile, they still face key challenges from two different perspectives: system usability and system efficiency.<br /> From the system usability perspective, current SWfMS are not designed to be simple enough for scientists who have quite limited IT knowledge. What’s more, there is no easy mechanism by which scientists can share and re-use scientific experiments that have already been designed and proved by others.<br /> From the perspective of system efficiency, existing SWfMS are coordinating and executing workflows in a centralized fashion using a single scheduler and / or a workflow enactor. This creates a single point of failure, forms a scalability bottleneck, and enforces centralized fault handling. In addition, they don’t consider load balancing while mapping abstract jobs onto several computational nodes. Another important challenge exists due to the common nature of scientific workflow applications, that need to exchange a huge amount of data during the execution process. Some available SWfMS use a mediator-based approach for data transfer where data must be transferred first to a centralized data manager, which is completely inefficient. Other SWfMS apply a peer-to-peer approach via data references. Even this approach is not sufficient for scientific workflows as a single complex scientific activity can produce an extensive amount of data.<br /> In this thesis, we introduce SWIMS (Scientific Workflow Integration and Management System) framework. It employs the Web Services technology to originate a distributed management system for data-intensive scientific workflows. The purpose of SWIMS is to overcome the previously mentioned challenges through a set of salient features: i) Support for distributed execution and management of workflows, ii) diminution of communication traffic, iii) support for smart re-run, iv) distributed fault handling and load balancing, v) ease of use, and vi) extensive sharing of scientific workflows. We discuss the motivation, design, and implementation of the SWIMS framework. Then, we evaluate it through the Montage application from the astronomy domain.