The cost abstraction for performance prediction will be conducted for each primitive operation. For computation components, an average cost function will be estimated using the parameters of the processor and cache/memory. For example, BLAS and LAPACK primitive performance cost functions will be studied in detail, and Rutgers/UCSB has done some preliminary studies on this topic. For standard communication and I/O primitives, cost functions will be determined based on device/networking performance parameters and the size of data communicated.
We will model the data communication structure between the modules (of Figure 2) seen in multidisciplinary and distributed applications. This coarse grain structure is typically not too sensitive to architectural details and we do not expect major difficulties in that area.