1. Introduction

An effort was initiated recently at Los Alamos National Laboratory (LANL) to build a new 3-D high-resolution tool for simulating casting processes, i.e. the flow of molten material into molds and the subsequent cooling and solidification of the material. The simulation process includes incompressible free-surface flow during mold filling, heat transfer-driven convective flows during solidification, and interface physics such as surface tension and phase change, all in complex geometries. This tool is known as TELLURIDE, and is described more fully elsewhere in these proceedings [8].

Several decisions were made early in the design stages of TELLURIDE which initiated and drove development of JTPACK90.

• An unstructured-mesh finite-volume approach would be used to the complex geometries that would be modelled.

• Fortran 90 (F90) was chosen as the implementation language. Though this was a fairly risky decision at the time due to the relative scarcity of stable compilers, we felt that the advantages over alternatives warranted the risk. For example, F90 offers numerous syntactic improvements to Fortran 77(F77), some of which will be discussed later. Compared with C++, we like the fact that arrays are first-class objects in F90, whereas in C++ every code effort seems to have its own array class. Also, the F90 standard had been approved and was in effect. Though this is not as much of a problem with C++ as it was when this effort was started, the standard still has not been finalized. We are now confident that our decision was the correct one, as F90 compilers have become available for virtually all platforms, and the syntactic advantages of F90 have, among other things, allowed us to write code which is maintainable and easy for new members to the team to become productive with.

• Parallelism would be via explicit message-passing using MPI (Message Passing Interface). In addition, the message-passing functionality would be encapsulated in an F90-accessible library, thus hiding the details from the higher-level code. This library is known as PGSLIB, and is described elsewhere in these proceedings [2].

• Robust and efficient solution of large systems of linear equations would be essential, as they would arise in several aspects of a simulation (e.g. heat conduction, our projection method for the Navier-Stokes equations, etc.).

One of us^e had developed an F77 package, JTPACK77 [10], which implements a number of Krylov subspace methods for solving linear systems of equations. Since that package was being used successfully for a number of efforts internal to LANL, it made sense to consider it as a candidate from which to build a similar F90 package for the TELLURIDE effort (and eventually other applications). In addition, although there are numerous other high-quality packages for iteratively solving systems of linear equations in other languages, such as ITPACK [7] and NSPCG [6] in F77, AZTEC [4] and PETSc [1] in C, and Diffpack in C++, we knew of no such effort in F90. So development of JTPACK90 began with JTPACK77 as a starting point, driven by the needs of TELLURIDE.