Grid'5000 experiment

Jump to: navigation, search

Finding best block size for out-of-core situation with optimized data representation (Programming)

Conducted by

Pierre-Nicolas Clauss


When using the optimized data layout it must be determined if there is still a best block size (that is: which gives the quickest results). The algorithm used for this experiment is the Livermore Loop Kernel 23. The goal is also to compare shared- and distributed- architectures with the same program, build through a common library.




  • Nodes involved: 10
  • Sites involved: 1
  • Minimum walltime: 24h
  • Batch mode: yes
  • Use kadeploy: no
  • CPU bound: yes
  • Memory bound: no
  • Storage bound: yes
  • Network bound: no
  • Interlink bound: no

Tools used

Tools written with parXXL (see


Results show that there is a block size that performs quicker results than any other block size (this results is not new, but was not shown for the optimized data layout yet). This experiment was not yet tested in distributed environment

Shared by: Pierre-Nicolas Clauss
Last update: 2010-03-14 01:25:54
Experiment #207