[an error occurred while processing this directive]

NERSC 3 Greenbook

Next: Access site: Up: Target for Energy Research Previous: Target for Energy Research

Local sites:

Most small sites, DOE University contracts, might consist of 1-3 workstations as described above connected to the Internet of the year 2000. In a good moderately-sized local installation, the workstations described above will be networked together and with more numerous smaller workstations by a gigabit/s local area network. These sites will have one or more 20-40 gigaFLOPS local compute servers and a nusers/resources/information server with a dozen or so terabytes on the local gigabit network. By 2000, some local systems will have a dynamic scheduler which will allow sharing the local compute servers and the background capacity of the workstations on the local area network. These schedulers might be able to schedule jobs that exceed the capacity of the local system on other systems in the ER organization. As developed below, ER should have a single large massively parallel system at an ``access site'' with the capability of a teraFLOPS or more with a memory of a teraword and a mass storage system of 100 terabytes. Software development and debugging would be done locally in the object-oriented environment of the workstations where the volume of the workstation market insures robust systems software and tools - not at the access site. Most local compute servers will be top-end symmetric multiprocessor systems available from the workstation manufacturers and will provide the same robust software environment. Thus the local compute servers would provide the resources for prototyping the large jobs that require the access center capabilities as well as providing much of the production (capacity) computing.

More specifically, computational scientists could expect a peak of 3 or more gigaFLOPS at night and more than 3 terabytes of mass storage from a typical local installation of ~4 large workstations and ~10 smaller workstations. If a compute server is local, the gaggle of workstations would most likely be larger and the site would have peak compute power of about 30-50 gigaFLOPS and have a mass storage system with 30-50 gigabytes. Most code and data would be maintained locally and only transported to the access site when needed there. (The large mass storage at the access site would be used to provide flexibility in the overall system, temporary storage for the input and output of calculations on the massively parallel system and as a repository of data used more than once by the high performance system.) More than 90% of all cycles used would be satisfied by local capacities. For example, most discrete calculations, simulations of a million or less particles, and most commercial applications requiring less than a few gigaFLOPS would be prevalent at the local sites. Local sites would have good graphics/visualization capabilities, possibly even virtual reality CAVEs, and other information infrastructural resources that need to be in the proximity of the users.

NERSC 3 Greenbook

Next: Access site: Up: Target for Energy Research Previous: Target for Energy Research