High Performance Computing Design by Code Migration for Distributed Desktop Computing Grids

Introduction

Grid system is a type of distributed systems, and it can be categorized into three classes: computing grid, data grid and service grid (Krauter et al., 2002; Choi et al., 2008). This article deals with the computing grid. The purpose of the computing grid is to use the existing computing powers effectively, those scattered around in one or more organizations. Several hundred PCs, such as Note PCs, Desktop PCs and WSs, are assumed in our model. User jobs, namely transactions, can be executed on the local or remote computer systems depending upon the workload in the system. If the workload of the system that the job entered is heavy enough for processing, it must be properly migrated to the other site through migration. However, the effectiveness of the code migration against the migration overhead must be evaluated (Coulouris et al., 2005).

The network of the computing grids we modeled is shown in Figure 1. It consists of many computing nodes and one information server. On decentralized control, it does not have the information server. Figure 2 shows the taxonomy of the distributed computing systems (Venugopal et al., 2006; Theotokis & Spinellis, 2004). The shadow area is the topics of this article. For migrating the transactions, both the transfer policy that determines the migration decision and the location policy that determines the site to be migrated must be properly provided according to the application environment (Shah et al., 2007). The environment of the system is getting more complex with the increasing and diversifying applications. However, little work for the code migration has been associated with the application environment. Depending upon the migration policy and the environment situated, the performance of the system greatly varies. The scales of transactions and networks, the workload patterns, the power of computers, code migration algorithms, all of these characterize the performance of the system. The relations of these characteristics, the pros and cons, must be analyzed, and the boundaries of the code migration must be specified in designing the distributed computing grids.

Figure 1.

Computing grid model

Figure 2.

Taxonomy of distributed computing systems

We need some methodologies to design the workload of transactions in various complex application environments to have a good performance system. Shah (2007) has shown the theoretical approach for code migration. We develop the practical approach in this article. Code migration overhead was exemplified by implementing the code migration system. Based on the experiments, several centralized and decentralized algorithms for various location policies are simulated and evaluated. And, design methodologies for distributed desktop computing grids in various environments are derived from several simulation results. The language for distributed computing was also designed to carry out the methodologies derived.

Experimental Model

The experimental system that migrates objects among PCs was implemented to measure the migration overhead. The system structure is shown in Figure 3. A transaction can be transferred to the other site to perform the computation by the agent. The performance of the two systems, the one that was executed at the local site without migrating object code and the other that was executed at the remote site, was measured and compared (Yoshida & Sakamoto, 2007).

Figure 3.

System structure