Grid'5000 experiment

Jump to: navigation, search

Heirarchal Data Partitioning For Linear Algebra Kernels on Clusters of Clusters (Other)

Conducted by

Brett Becker


We are currently working on a new heirarchal data partitioning scheme for linear algebra kernels on clusters of clusters. Our primary focus is on minimising the amount of inter-cluster communication. Other avenues include overlapping communication and computation and expanding the data partitioning scheme to other architectures and problem domains.


in progress


  • Nodes involved: 100
  • Sites involved: 3
  • Minimum walltime: 1h
  • Batch mode: yes
  • Use kadeploy: no
  • CPU bound: yes
  • Memory bound: no
  • Storage bound: no
  • Network bound: no
  • Interlink bound: no

Tools used

Open-Mpi, Infiniband


A new partitioning algorithms for matrix multiplication and related operations designed for use on a small number (2, 3, ...) of clusters/sites has been developed and demonstrated which minimizes the total volume of inter-cluster communication optimally, and better than all other known data partitioning algorithms.

Shared by: Brett Becker
Last update: 2011-04-12 23:56:26
Experiment #529