Extensible job managers for grid computing

Abstract

Grid computing is becoming an important framework for enabling applications to utilize widely distributed collections of computational and data resources, however current grid software is still immature and rather difficult to use. The Globus Grid Toolkit is a set of low-level tools, protocols and services that has become a defacto standard for basic grid computing infrastructure. The Globus Resource Allocation and Management (GRAM) service provides for the management and remote execution of jobs defined using a standard Resource Specification Language (RSL). Currently, the GRAM has very limited functionality, which makes it more difficult to develop grid applications. One limitation is the lack of support for applications that require a special execution environment, such as Java applications that run within a Java Virtual Machine. Cumbersome workarounds are necessary to run such applications. The current GRAM addresses these problems in a rather ad hoc way for certain specific cases, however there is no general, well- defined mechanism for supporting arbitrary execution environments. Here we outline some of the problems with the current Globus GRAM specification and provide a proposal for how they might be addressed by defining some extensions to the standard RSL supported by the GRAM, as well as some modifications to the design of the GRAM that would enable it to support arbitrary execution environments. We give examples of how our proposed system can provide improved support for Java applications and cluster management systems, and describe our ongoing work in implementing prototypes of these proposed GRAM extensions.