Grid-enabling Non-computer Resources – Københavns Universitet

Grid-enabling Non-computer Resources

PhD thesis defence by Martin Rehr

Abstract


This dissertation addresses the problem of harvesting user-contributed computer resources that may be constrained beyond a conventional computer for Grid computing. In order to make users donate their computing power to a Grid computing infrastructure, minimum intrusion of the donated resource is crucial. That is, the administrative task of connecting the resource to the infrastructure should be kept at a minimum and security should be kept at a maximum ensuring that the donated machine can not be harmed.

The first part of the dissertation covers which mechanisms are needed to accomplish minimum intrusion and have resulted in two models: "The One-Click Grid-resource model" and "The PS3 Grid-resource model". The first model covering how every Java enabled web-browser can be turned into a sandboxed Java Grid resource just by accessing an URL in a browser. The second covering how the Playstation 3 game console can be turned into a sandboxed High Performance Grid computing resource by booting it with a live-cd.

One problem regarding user donated Grid resources is the heterogeneity of such resources. Today CPUs are virtualized to make them appear homogeneous to Grid applications, disk space is provided through remote file access frameworks in order to make them virtual infinite, but yet there have been no way of providing virtual infinite memory to Grid applications.

The last part of the dissertation addresses this problem by introducing a transparent user-level remote memory framework. This framework is transparent both to the Grid application and the Grid resource, because it operates as a layer between the user application and the operating system. This means that the framework can be deployed through the Grid infrastructure along with the Grid jobs without human interaction or privileged access to the resource.

Experiments show that the remote user-level memory framework outperforms swap to local disk in a low-latency high bandwidth network and by using prediction based prefetching also outperform swap to local disk in a high-latency, high-bandwidth network.

Assessment Committee:


Chairman: Associated Professor Knud Henriksen, DIKU, Copenhagen
Member 1: Associate Professor John Markus Bjørndalen, Tromsø University, Norway
Member 2: Associated Professor Josva Kleist, Aalborg University, Denmark

For an electronic copy of the thesis, please contact Dina Riis Johannessen, dinariis@diku.dk