Amril Nazir

A Cost Ef?cient Framework for Managing Distributed Resources in a Cluster Environment

Most resource management systems in large distributed environments rely on a centralised controller scheme, known as a meta-scheduler. Such a meta-scheduler is designed to have a complete control of all resources worldwide through a number of local schedulers on each participating site. Faced with a potential large number of sites, the central scheduling point represents a potential bottleneck for requests needing fast response. In this paper, we consider the ef¿ciency of a localised two-tier system: the ¿rst tier selects appropriate computing resources from the worldwide pool and creates a virtual cluster. The second tier employs a local scheduler to distribute jobs/tasks from the applications among the resources that make up the cluster. The two-tier approach differentiates the tasks of securing a suitable resource level and managing their use. The aim is to ensure a rapid request response while at the same time optimising the system performance. To this end, we consider a management strategy that incorporates rental policy and scheduling schemes. We propose aggressive and conservative heuristic based policies and evaluate their performance to demonstrate the cost effectiveness of such approaches.

Cost-benefit analysis of high performance computing infrastructures

The Grid computing vision promises to provide the required platform for users to run a new and more demanding range of high performance computing (HPC) applications. However, the performance i.e. the response times achievable from the grid systems, is reported as not very satisfactory. End users may be forced to purchase, maintain and upgrade their own dedicated HPC systems for the execution of their applications, which could be a far more expensive solution. But is there an alternative option? To answer this question, this paper carries out total cost of ownership (TCO) analysis of a private resource system and dedicated large HPC system. In particular, we detail the specific costs of a private resource system from a LUNAR project, a large dedicated HPC system from the Lawrence Livermore National Lab, and international EGEE grid project. With this data, we compare the performance and cost-benefits of a dedicated HPC system in contrast to a private resource system. Our evaluation demonstrates that a small private resource system with renting capability can provide a viable alternative to users to run their HPC applications, without the need to purchase and maintain a large, dedicated, HPC cluster system.

A Software Framework to Support Adaptive Applications in Distributed/Parallel Computing

Resource allocations are performed statically for traditional distributed/parallel applications prior to launching application executions. This limitation could cause long resource waiting time when multiple resources need to be co-allocated. In contrast, the Adaptive Distributed/Parallel Applications (ADA), allowing resources to be added and released during executions, can adapt to the dynamic feature of common distributed computing environments (e.g. the Grid). The goal of the work is to create a tool to allow users to easily develop and run ADAs without dealing with the underneath distributed resource environments. We introduce a novel software package, Application Agent (AA), to support the execution of ADAs, including automatic resource allocation, dynamic process deployment, and enabling process wide-area communication. An AA-enabled application can be started on any internet-connected machines and the AA will dynamically configure a virtual machine from the local machine to remote available machines to satisfy the execution. The AA is composed of two parts. The first part is a library of AA interface routines, which contains user-callable functions for developers to integrate their applications with the AA. This second part is the daemons, which dynamically collect computational resources on the Internet to create a wide-area virtual machine to execute an application.

Preliminary Resource Management for Dynamic Parallel Applications in the Grid

Dynamic parallel applications such as CFD-OG impose a new problem for distributed processing because of their dynamic resource requirements at run-time. These applications are difficult to adapt in the current distributed processing model (such as the Grid) due to a lack of interface for them to directly communicate with the runtime system and the delay of resource allocation. In this paper, we propose a novel mechanism, the Application Agent (AA) embedded between an application and the underlying conventional Grid middleware to support the dynamic resource requests on the fly. We introduce AA’s dynamic process management functionality and its resource buffer policies which efficiently store resources in advance to maintain the execution performance of the application. To this end, we introduce the implementation of AA.

Service Level Agreements in a Rental-based System

In this paper, we investigate how Service Level Agreeements (SLAs) can be incorporated as part of the system’s scheduling and rental decisions to satisfy the different performance promises of high performance computing (HPC) applications. Such SLAs are contracts which specify a set of application-driven requirements such as the estimated total load, contract duration, total utility value and the estimated total number of generated jobs. We present several scheduling and rental based policies that make use of these SLA parameters and demonstrate the effectiveness of such policies to accurately predict and plan for resource levels in a rental-based system.

A Rental-Based Approach in a Cluster or a Grid Environment

We examine the benefits of a rental based cluster system in supporting high-performance applications. We evaluate the cost of five rental policies and demonstrate how resources from third party external resource providers can be efficiently obtained to accommodate sudden bursts in demand for resources by high performance computing applications. Our evaluation demonstrated that the proposed policies can achieve a balance between performance and usage cost, and show how they are able to attain higher profit by lowering both penalty and rental costs.

Supporting Dynamic Parallel Application Execution in the Grid

Dynamic parallel applications such as CFD-OG impose a new problem for distributed processing due to their dynamic resource requirements at run-time. These applications are difficult to adapt in the current distributed processing model (such as the Grid) because of a lack of interface for them to directly communicate with the runtime systemand the delay of resource allocation. In this paper, we propose a novel mechanism, the Application Agent (AA) embedded between an application and the underlying conventional Grid middleware to support the dynamic resource requests on the fly. We introduce AAs dynamic process management functionality and its resource buffer policies which stores resources in advance to maintain the execution performance of the application.

Resource Management Support for Smooth Application Execution in a Dynamic and Competitive Environment

Under a competitive grid environment, the time delay involved in assigning resources can ruin the smooth execution of a highly interactive application with dynamic resource requirements. To solve this problem, we propose two novel approaches to be embedded on top of the generic resource management. One approach is the priority-based allocation policy that gives advantages for the urgent requests; another is the resource permit swapping policy that is proposed to resolve resource requirements conflicts. Experimental results and analyses demonstrate the effectiveness of our approach in terms of the satisfaction of urgent requests in maintaining the smooth time progression for highly dynamic applications and also the satisfaction of other types of applications.