A comparison between the sequential and parallel SIMPDEL models was made based on a 23 year simulation for each initial population size. Execution times and speed improvements are given in Table 1.
The sequential program was executed on a Sun SPARCstation 5, with 32 Mbytes of memory and 640 Mbytes of disk space. The parallel program was executed on a 32 processor Thinking Machines CM-5 with 32 Mbytes of memory on each SPARC 2 processor. Speed improvements were calculated using wall-clock times; however, cpu times for the sequential model are also presented (Table 2).
Execution times were recorded for the initial deer population sizes: 2,000, 10,000, and 20,000, with a peak speed improvement of 27 for the parallel model over the sequential model for the population size of 2,000. Speed improvements for the hydrology and vegetation components are larger than the number of processors, due to the greater memory limitations of the sequential computing environment. For example, execution of the sequential model produces approximately 100,000 page faults per simulation year for an initial population size of 10,000, however with 32 Mbytes of memory on each processor of the CM-5, all data for the parallel model fits in main memory. In addition, memory requirements are greater in the sequential model since about 1/3 more map data is stored than in the parallel model.
As the sequential model is updated and improved, speed improvements in the hydrology and vegetation components will most likely decrease. By simply replacing the arrays in the sequential model with a map data structure similar to that of the parallel model, execution times should decrease somewhat. The parallel map data structure has two main advantages over the various arrays used to store map data in the sequential model. First, the data structure enables the storage of only the valid grid cells (those representing grid cells within the study area) and provides a means of mapping each grid cell onto its actual position in the landscape. Second, all map data corresponding to the same 500m grid cell is encapsulated into one structure and stored in contiguous memory. Since most computations require the use of data values located in the same array index, but from several different arrays, the parallel map data structure provides faster access to data values.