Implementing Production Grids

We have presented the essence of experience gained in building two production Grids, and provided some of the global context for this work. As the reader might imagine, there were a lot of false starts, refinements to the approaches and to the software, and several substantial integration projects (SRB and Condor integrated with Globus) to get where we are today. However, the point of this paper is to try and make it substantially easier for others to get to the point where Information Power Grids (IPG) and the DOE Science Grids are today. This is what is needed in order to move us toward the vision of a common cyber infrastructure for science. The author would also like to remind the readers that this paper primarily represents the actual experiences that resulted from specific architectural and software choices during the design and implementation of these two Grids. The choices made were dictated by the criteria laid out in section 1. There is a lot more Grid software available today that there was four years ago, and various of these packages are being integrated into IPG and the DOE Grids. However, the foundation choices of Globus, SRB, and Condor would not be significantly different today than they were four years ago. Nonetheless, if the GGF is successful in its work - and we have every reason to believe that it will be - then in a few years we will see that the 28 functions provided by these packages will be defined in terms of protocols and MIS, and there will be several robust implementations available for each of the basic components, especially the Grid Common Services. The impact of the emerging Web Grid Services work is not yet clear. It will likely have a substantial impact on building higher level services, however it is the opinion of the author that this will in no way obviate the need for the Grid Common Services. These are the foundation of Grids, and the focus of almost all of the operational and persistent infrastructure aspects of Grids.