Acyclic Partitioning of Large Directed Acyclic Graphs - Inria - Institut national de recherche en sciences et technologies du numérique
Communication Dans Un Congrès Année : 2017

Acyclic Partitioning of Large Directed Acyclic Graphs

Résumé

Finding a good partition of a computational directed acyclic graph associated with an algorithm can help find an execution pattern improving data locality, conduct an analysis of data movement, and expose parallel steps. The partition is required to be acyclic, i.e., the inter-part edges between the vertices from different parts should preserve an acyclic dependency structure among the parts. In this work, we adopt the multilevel approach with coarsening, initial partitioning, and refinement phases for acyclic partitioning of directed acyclic graphs and develop a direct k-way partitioning scheme. To the best of our knowledge, no such scheme exists in the literature. To ensure the acyclicity of the partition at all times, we propose novel and efficient coarsening and refinement heuristics. The quality of the computed acyclic partitions is assessed by computing the edge cut, the total volume of communication between the parts, and the critical path latencies. We use the solution returned by well-known undirected graph partitioners as a baseline to evaluate our acyclic partitioner, knowing that the space of solution is more restricted in our problem. The experiments are run on large graphs arising from linear algebra applications.
Fichier principal
Vignette du fichier
paper.pdf (467.63 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Loading...

Dates et versions

hal-01672010 , version 1 (22-12-2017)

Identifiants

Citer

Julien Herrmann, Jonathan Kho, Bora Uçar, Kamer Kaya, Ümit V. Çatalyürek. Acyclic Partitioning of Large Directed Acyclic Graphs. CCGRID 2017 - 17th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing, May 2017, Madrid, Spain. pp.371-380, ⟨10.1109/CCGRID.2017.101⟩. ⟨hal-01672010⟩
255 Consultations
3449 Téléchargements

Altmetric

Partager

More