It has often been suggested by managers and others that NERSC cannot be a place for ``midlevel'' computing, or ``routine'' computing, and that such computing should be moved elsewhere. We believe the distinctions of ``midlevel'' or ``routine'' vs. ``true'' supercomputing to be largely improper and artificial. Something that a manager might consider ``midlevel'' may be viewed by the user as a necessary part of his entire supercomputing endeavor, and one that requires substantial computing capacity. When there are no new resources, some hard judgments have to made about allocating existing resources. But the computer industry has been making big strides in the advancement of lower price/performance platforms, particularly with the now available SMPs. Acquisition of such platforms is evidently cost-effective. The stumbling block appears to be the policies in Washington, over two particular issues:
Big SMPs with big memories not only provide a cost-effective route to badly needed increased capacity, but can also provide high capability compared to the kind of workstation resources that are typically available at local sites. Another extremely favorable point for SMPs: for many of our codes there is no large new programming effort needed. As development platforms, the SMPs would directly contribute to the best use of the MPP, for which it is more difficult to develop software. We expect development on an SMP followed by production on the MPP to be the most successful transition path to effective use of the MPP. We urge that high priority be given to the acquisition of some of these cost-effective and versatile platforms.