The specialized data management system described in this paper was motivated by the need for much more efficient data management than a standard database management system could provide for particle physics codes in shared memory multiprocessor environments. The special characteristics of data and access patterns in particle physics codes need to be fully exploited in order to effect efficient data management. The data management system allows parameteric user control over system features not usually available to them, especially details of physical design and retrieval such as horizontal clustering, asynchronous I/O, and automatic distribution across processors. In the past, each physics code has constructed the equivalent of a primitive data management system from scratch. The system described in this paper is a generic system that can now be interfaced with a variety of physics codes.

Current research into data modelling of large scientific simulation programs is described. The research includes an empirical study of data management in such programs and an analysis of how scientists actually use their data. The goal of the research is to provide theoretical tools to improve data management for this class of applications. In this paper, data modelling in the scientific environment is described, and some comparisons between data modelling for scientific simulation programs and commercial applications are made.