Fig 1.
Parallelization of Dust Model (Domain Division and Two Allocation Methods for Dispatching 12 Subdomains to Two Computing Nodes, A and B).
Fig 2.
Workflow of K-Means & Kernighan-Lin Combined Algorithm.
Fig 3.
Subdomain Center Points and Component Centroids within a Domain Matrix.
Fig 4.
Workflow of Subdomain Belongingness Adjustment.
Fig 5.
Pseudo codes of Algorithm SwapSubdomains.
Fig 6.
Workflow of ASRG algorithm.
Fig 7.
Implementation of ASRG algorithm.
Table 1.
Scenario Description.
Table 2.
Programming language and used software/libraries for algorithms.
Fig 8.
Total Shared Edge Number Comparison for Different Domain Sizes.
Table 3.
The difference of the total number of shared edges between K&K and ASRG.
Fig 9.
Standard Deviation of the Number of Subdomains in Each Components Comparison for Different Domain Sizes.
Table 4.
The maximum difference on the number of subdomains between K&K and ideal algorithm.
Fig 10.
Standard Deviation of the Number of Shared Edges in Each Components Comparison for Different Domain Sizes.
Table 5.
The maximum difference of the number of shared edges for K&K and ASRG.
Fig 11.
Solving Time Comparison for Different Domain Sizes.
Fig 12.
ILP Solving Time Growth with the Number of the Components and the Domain Size (the solving processes were force to be terminated when taking longer than 7200s).
Table 6.
Comparison of three algorithms through different features.
Fig 13.
Subdomain Number—Execution Time Plot for Different Node Numbers (Blue bars: execution time using MPI default allocation; Orange bars: execution time using K-Means and Kernighan-Lin combined algorithm (K&K); Grey lines: Performance Improvement Factor (PIF)).
Fig 14.
Node Number—Execution Time Plot for Different Subdomain Numbers (Blue bars: execution time using MPI default allocation; Orange bars: execution time using K-Means and Kernighan-Lin combined algorithm (K&K); Grey lines: Performance Improvement Factor (PIF)).
Fig 15.
PIF and Estimated Communication Volume for four Nodes.
(A) shows the PIF values for 4 nodes. (B) and (D) are the allocation conFig uration using K&K method and default method. (C) is the estimated communication volume (both external and internal) for 4 nodes using two methods. Blue bars: estimated external communication volume using MPI default allocation; Orange bars: estimated internal communication volume using MPI default allocation; Grey bars: estimated external communication volume using K&K allocation; Yellow bars: estimated internal communication volume using K&K allocation.