Overview and evaluation of conceptual strategies for accessing CPU-dependent execution resources in grid infrastructures

Abstract

The emergence of many-core and massively-parallel computational accelerators (e.g., GPGPUs) has led to user demand for such resources in grid infrastructures. A widely adopted approach for discovering and accessing such resources has, however, yet to emerge. GPGPUs are an example of a larger class of computational resources, characterized in part by dependence on an allocated CPU. This paper terms such resources »CPU-Dependent Execution Resources« (CDERs). Five conceptual strategies for discovering and accessing CDERs are described and evaluated against key criteria, and all five strategies are compliant with GLUE 1.3, GLUE 2.0, or both. From this evaluation, two of the presented strategies clearly emerge as providing the greatest flexibility for publishing both static and dynamic CDER information and identifying CDERs that satisfy specific job requirements. Furthermore, a two-phase approach to job-submission is proposed for those jobs requiring access to CDERs. The approach is compatible with existing grid services. Examples are provided to illustrate job submission under each strategy.