We close with three points and a final statement.
One, the problems the community is experiencing are a consequence of our success. Overcrowding occurs because NERSC is successful at providing a unified computing environment to a broad base of users. And the decrease in general access caused by the sudden jump in SPP only occurred because of the success of this new way to use the C90: computations heretofore impossible are now routine in the SPP program, extending the scope of research to new regimes.
Two, the most important product of NERSC is the new scientific research that comes from the use of the computing resources. If this product is not continually pushed towards excellence, we shouldn't be in business at all.
Three, we believe that Washington has an opportunity to renew its support for the model of broad based supercomputer access. This includes extending the high end at the pace available from the computer industry (via MPP) and supporting its very broad base of users at the same time (via increasingly cost-effective SMP platforms now available). For us to continue to adhere to policies that actively advance the cause of only the highest-end users will only ensure the slow decay of NERSC; computational scientific research throughout ER will be the loser.
A diverse group of OER researchers uses NERSC to produce high-quality scientific results. It is now time to take advantage of the enormous advances made by the American computer industry in increasing the compute cycle/dollar ratio. The near-future acquisition of an MPP is seen as the best way to enhance the NERSC computing resources to push science beyond its current scope. In the meantime the addition of SMP systems will improve the facility by: (1) providing a development platform for the labor intensive process of converting research applications to use MPPs, (2) allowing the expansion of the SPP program to allow an increased subset of users the ability to do new science, and (3) providing a capacity engine for appropriate applications and thus reducing the load on the current vector supercomputers.