Performance Measurement of Scientific Grid-Roadmap to eScience

Abstract

The term grid was rst proposed in mid-90's and in short span of time has attracted lot of attention from researchers in computer science and its applications. The growth of grid and its usage has also created challenges for computer scientists to make it more reliable and powerful tool in hands of researchers in every discipline who are faced with large scale computing problems. However the development of compute intensive ap- plication is in practice, very di cult and time consuming, because of the need to deal with complex & heterogeneous system. This thesis aims at demonstrating the research work carried out speci cally on computational grid and enhancing its performance by designing and implementing algorithms which solves some issues. Carrying this inten- tion, user driven interactive scheduler is also developed compatible with Globus, which is de-facto middleware standard for grid. To corroborate the grid functionality, A - testbed developed on University Campus, hereafter called NCampusGrid is engineered from idle desktop of university campus in which cycle harvesting is attained by suit of grid and middleware technologies such as Globus,Condor,PBS and SGE, further optimized to extract large computational ca- pacity in the form of MIPS and GFLOPS. This heterogeneous grid architecture has proved to be basic axiom of grid which is often regarded as a way for creating inex- pensive \super-computing". Desiring to achieve \performance" as a basic motivation of Computational grid, scienti c problems are deployed on grid requiring a high level of expertise and a sig- ni cant amount of e orts, mainly due to overall complexity of the grid. Initial part of this research reveals the grid enabled computing capabilities for climate impact modeling and change detection facilitating for rainfall prediction. The large dataset utilized in computations are Hamburg Ocean Atmosphere Parameters and Fluxes from Satellite (HOAPS) data presenting the basic eld of air-sea interaction parameters and covers the period 1987 to 2005. The grid enabled application designed, developed and deployed on NCampusGrid testbed, provides user friendly interface, secured job submission for heterogeneous environment needed computational power required for analyzing large amount of scienti c data. E orts are made not just to measure the performance of computational grid but also to enhance its performance. While dealing with computational grid, an attempt has been made to achieve load balancing, optimum resource utilization and reliability etc. Most of the grid middleware supports batch processing of jobs which is a bottle- neck for using grid for interactive jobs, or provide user friendly environment. In current situation starting from the console based job submission to monitoring, the process is cumbersome. Speci cally in long running job, lack of interactive support creates dif- culty in tracing the failure of job which may sometime lead to wastage of resources. Overcoming these lacunas, e orts taken for this project have resulted into development of Globus compatible interactive user driven scheduler, called ISched, facilitating user in managing the local and remote jobs, as Globus uses third party scheduler. In the analysis and design of the scheduler for interactive jobs, formal methods have been followed. The scheduler is designed to balance the load under the user de ned constraints.