next up previous
Next: Local Temporary Files (LTF) Up: Parallel Algorithms Previous: Parallel Algorithms

Global Temporary Files (GTF)

The first algorithm we describe relies on global temporary databases for the selected lines. This is the algorithm that was implemented in the versions of PHOENIX discussed in papers I and II. In the general case of N available PE's, the parallel line selection algorithm uses one PE dedicated to I/O and (N-1) line selection PEs. The I/O PE receives data for the selected lines from the line selection PEs, assembles them into properly sorted blocks of selected lines, and writes them into the temporary database for later retrieval. The (N-1) line selection PEs each read one block from a set of (N-1) adjacent blocks of line data from the master database, select the relevant lines, and send the necessary data to the I/O PE. Each line selection PE will select a different number of lines, so the I/O PE has to perform administrative work to construct sorted blocks of selected lines that it then writes into the temporary database. The block sizes for the line selection PEs and for the temporary database created by the I/O PE do not have to be equal, but can be chosen for convenience. The blocks of the master line database are distributed to the (N-1) line selection PEs in a round robin fashion. Statistically this results in a balanced load between the line selection PEs due to the physical properties of the line data.

After the line selection phase is completed, the temporary global line database is used in the line opacity calculations. If each on the N PEs is calculating line opacities (potentially for different sets of wavelengths points or for different sets of physical conditions), they all access the temporary database simultaneously, reading blocks of line data as required. In most cases of practical interest, the same block of line data will be accessed by several (all) PEs at roughly the same time. This can be advantageous or problematic, depending on the structure of the file system on which the database resides. The PEs also cache files locally (both through the operating system and in the code itself through internal buffers) to reduce explicit disk I/O. Note that during the line selection phase the temporary database is a write-only file, whereas during the opacity calculations the temporary database is strictly read-only.

The performance of the GTF algorithm depends strongly on the performance of the global file system used to store the temporary databases and on the characteristics of the individual PEs. This issue is discussed in more detail below.


next up previous
Next: Local Temporary Files (LTF) Up: Parallel Algorithms Previous: Parallel Algorithms
Peter Hauschildt
2001-04-16