Scalability of Sparse Solvers in the Context of Semiconductor Device Simulation Applications

In this paper, we present our preliminary experience in parallelizing
IBM's FIELDAY code which is a general purpose finite element code for
simulating two- and three-dimensional semiconductor devices using the
conventional drift-diffusion equations. We discuss some of the key
features of our implementation, focusing primarily on the handling of
the irregular grid computations and on the solution of the linear
systems. In this implementation, the linearized systems are solved by
preconditioned (using incomplete LU) conjugate gradient squared (CGS)
type of iterative methods. Our preliminary performance data indicates
that while scalability on a per iteration basis is within acceptable
range for sparse, irregular problems, the overall scalability is poor.
Our experiments so far indicate that the primary reason for the
overall poor performance is the inherent tradeoff seen between the
available parallelism and the convergence rate of the iterative
methods used. By making the grid partitioner aware of the underlying
physical processes involved in the device being simulated, we hope to
improve the convergence behavior without compromising on the
extractable parallelism.