Introduction

In the first 2 papers of this series we have described general parallel algorithms that we have implemented in our multi-purpose stellar atmosphere code PHOENIX. These papers focused mainly on radiative transfer and NLTE problems and general parallelization issues. In this paper we discuss in greater depth the problem of line opacity calculations. This is in particular important if extensive line lists are used, i.e., molecular line databases. These databases have increased dramatically over the last decade, mostly due to the work of on water vapor lines and on TiO lines. Currently our molecular line database contains about 550 million lines, most of which are TiO and H₂O lines. Using these databases for opacity calculations poses a significant challenge both for the construction of opacity tables and for the construction of detailed model atmospheres.

In our model atmosphere code we have implemented and used direct opacity sampling (dOS) for more than a decade with very good results. During that time, the size of the combined atomic and molecular line databases that we used has increased from a few MB to $>8\,$GB. Whereas the floating point and memory performance of computers has increased dramatically in this time, I/O performance has not kept up with this speed increase. Presently, the wallclock times used by the line selection and opacity modules are dominated by I/O time, not by floating point or overall CPU performance. Therefore, I/O performance is today more important that it was 10 years ago and has to be considered a major issue. The availability of large scale parallel supercomputers that have effectively replaced vector machines in the last 5 years, has opened up a number of opportunities for improvements of dOS algorithms. Parallel dOS algorithms with an emphasis on the handling of large molecular line databases are thus an important problem in computational stellar atmospheres. These algorithms have to be portable and should perform well for different types of parallel machines, from cheap PC clusters using Ethernet links to high performance parallel supercomputers. This goal is extremely hard to attain on all these different systems, and we, therefore, consider two different parallel dOS algorithms in this paper and compare their performance on two very different parallel machines. In the next sections we will discuss the direct opacity sampling method, describe the parallel algorithms in detail and then discuss the results of test calculations. We close with a summary and conclusions.

on TiO lines. Currently our molecular line database contains about 550 million lines, most of which are TiO and H₂O lines. Using these databases for opacity calculations poses a significant challenge both for the construction of opacity tables and for the construction of detailed model atmospheres.

In our model atmosphere code we have implemented and used direct opacity sampling (dOS) for more than a decade with very good results. During that time, the size of the combined atomic and molecular line databases that we used has increased from a few MB to $>8\,$GB. Whereas the floating point and memory performance of computers has increased dramatically in this time, I/O performance has not kept up with this speed increase. Presently, the wallclock times used by the line selection and opacity modules are dominated by I/O time, not by floating point or overall CPU performance. Therefore, I/O performance is today more important that it was 10 years ago and has to be considered a major issue. The availability of large scale parallel supercomputers that have effectively replaced vector machines in the last 5 years, has opened up a number of opportunities for improvements of dOS algorithms. Parallel dOS algorithms with an emphasis on the handling of large molecular line databases are thus an important problem in computational stellar atmospheres. These algorithms have to be portable and should perform well for different types of parallel machines, from cheap PC clusters using Ethernet links to high performance parallel supercomputers. This goal is extremely hard to attain on all these different systems, and we, therefore, consider two different parallel dOS algorithms in this paper and compare their performance on two very different parallel machines. In the next sections we will discuss the direct opacity sampling method, describe the parallel algorithms in detail and then discuss the results of test calculations. We close with a summary and conclusions.