Computational modeling and simulation is now accepted  as a  common
mode for knowledge discovery and design across science, engineering
and manufacturing.  In many such applications, the models are  based
on partial-differential equations which are typically solved
numerically using sparse or irregular scientific computing kernels.
Such sparse kernels differ greatly in their  memory, cache and FPU
usage  from  the SPEC floating-point benchmark kernels and the more
traditional dense matrix LINPACK benchmark.  For example, dense
matrix kernels have high levels of data locality and re-use that
enable them to execute at near peak rates on high performance
microprocessor architectures with deep cache hierarchies and high
clock frequencies.  However, sparse kernels can utilize only  a
fraction of the available processing speed because they  have a
large number of data accesses per floating-point operation, and
limited data locality  and data re-use despite algorithmic changes
and considerable tuning of codes through blocking and loop unrolling
schemes.

In this paper, we take into account current trends in power-aware
high performance architecture design to explore the impact of memory
subsystem optimizations on sparse scientific codes. We propose a
variety of features to enable energy-aware, high performance, sparse
scientific applications.  We use simulations with SimpleScalar and
Wattch to demonstrate the effectiveness of our approach for an
optimized sparse-matrix-vector multiplication kernel  and two sparse
kernels from the NAS benchmark.  Our results indicate that in certain
cases, our optimizations can improve cache miss rates by  90\%,
application time by 80%, and  the energy $\times$  time metric by 90\%,
at system power levels less than or equal to those in the original
configuration.