On the Pervasive Adoption of Grid Technologies: A Grid Operating System

Introduction

Despite having made substantial advances during the last decade, grid computing is still neither pervasive nor widely deployed. Gartner predicted that by 2006 grid computing would mature sufficiently to leave the science laboratories and enter the business world (Gartner Group, 2003). But so far there have been only a few success stories, since only a subset of business applications are supported by existing grid infrastructures. To date the computing research community and particularly eScience projects have been the biggest beneficiaries of grid computing whereas other communities, such as user-centric fields like medical sciences, cannot easily implement existing grid architectures to support their applications. In Mattmann (2007), various requirements have been outlined which existing grid middleware do not support, thus increasing the cost of the adoption of grids for medical science. It is a similar story for enterprise computing where existing grid middleware is not scalable or lacks fault tolerance. Moreover they are mostly platform dependent and are insufficiently flexible to support enterprise applications. These issues in grid adoption can be traced to technical hurdles which arise as a consequence of the current approaches to grid computing. We are of the opinion that these hurdles originate from the current middleware approach to grid computing, as detailed below.

The middleware approach to grid computing was developed in science laboratories in which computing clusters distributed across the world are linked together in order to create grids to solve mainly compute and data intensive scientific problems. The role of grid middleware in this paradigm was to ‘glue’ the clusters together to achieve interoperability. Notable grid middleware included Globus (Foster and Kesselman, 1997), gLite (Laure, 2004) and UNICORE (Erwin and Snelling, 2001). This approach has however created some obstacles to grid adoption in other fields, since the cluster-oriented grids of today are not very suitable for user-centric computation, due in part to their complex operation and maintenance requirements. The main barriers (Ali, 2006) to the adoption of grid computing that result from the strong focus of current grid computing research on eScience are: the support for only limited application types (which mostly comprise highly parallel and batch applications), potentially inflexible network topologies, the steep learning curve required for configuring and maintaining a grid with grid middleware, the lack of fault tolerance in the infrastructure and an inflexibility of virtual organization (VO) management software to create more fine grained VOs, as required by some applications. All of these limitations make grid computing in its present incarnation unsuitable for the common user with little computing expertise and make grid computing expensive for existing users to maintain.