Service oriented approach to high performance scientific computing

Abstract

Service Oriented Architectures (SOA) are used in business circles to harness the distributed resources of an enterprise, with relative ease and convenience. SOA is perceived to deliver programmability, scalability and efficiency under heterogeneity. The scientific community is always aspired to have a similar convenient approach to solve "not so embarrassingly parallel" scientific problems with expected levels of high performance, especially in heterogeneous conditions. One of the major challenges in parallelizing scientific algorithms is their apparent interdependency of tasks (atomic units of parallel works) which results in too fine granularity. The computational advantages in parallelizing those scientific algorithms can be overshadowed by costs involved in communications between non-optimal fine-grained tasks, in a SOA environment. The aim of this PhD research is to overcome these challenges and to empower scientists and researchers with SOA tools to develop high performance scientific applications with relative ease that can perform well under heterogeneous environments. The research has produced a scalable and heterogeneity-oblivious SOA middleware - ANU-SOAM. It implements a popular enterprise SOA middleware API (IBl\II-Platform Symphony API) and thus ensures programmability. It offers better performance under heterogeneous conditions by implementing load balancing and scheduling techniques. Along with its compute services it provides a Data Service which helps application programmers to develop codes that can effectively circumvent the interdependency of tasks and thereby reduce communications to ensure high performance outcomes. The Data Service achieves this by allowing data to be stored, accessed, modified and synchronized (using add, get, put and sync functionalities) at host and compute nodes according to the application logic. It is also observed that the programming model supported by the Data Service can help ANU-SOAM applications to access compute resources in a "Cloud IaaS" over high latency networks (like the Internet) with much lower overheads compared to the conventional SOA programming models.