Rank reordering for MPI communication optimization

Communication-aware process and thread mapping using online communication detection

“…Other algorithms have been proposed, such as Treematch [35] and Zoltan [36]. We selected Scotch because it has a short execution time (less than 1 ms to map 32 processes) while providing good results [37][38][39]13,40], and is therefore suitable for online mapping. Scotch has a complexity of OðN 3 Þ, where N is the number of processes to be mapped [38].…”

Recent Advances in Graph Partitioning

“…Greedy approaches such as the one by Brandfass et al [BAG13] map the node v c of G c with the highest total communication cost w. r. t. to the already mapped nodes onto the node v p of G p with the smallest total distance w. r. t. to the already mapped nodes. Some variations exist that improve this generic approach in certain settings [HS11,GMN15].…”

“…Brandfass et al [BAG13] present local search and evolutionary algorithms. Their experiments confirm that metaheuristics are significantly slower than problem-specific heuristics, but obtain highquality solutions [BAG13,UAKI06].…”

Communication and topology-aware load balancing in Charm++ with TreeMatch

“…MPI 2.2 is a relevant example [10]. Beyond a static placement, some works focus on online placement by relying on a technique called rank reordering ( [7], [4]). Finally, Dummler [11] explored the issue of hybrid, MPI + OpenMP application multithreaded process mapping.…”

“…This technique can improve application performance and ultimately scalability since communication costs are decreased. So far, a compelling idea is to try and regroup on the same computing node the processes sharing the most data so as to reduce the application communication costs [4], [5]. However, it now becomes necessary to push further this idea and apply it to the subsets of application processes within a computing node to take advantage of the complex structure of its memory.…”