parXXL: Scaling Fine Grained Algorithm Design to Grids

1
parXXL Scaling Fine Grained Algorithm Design
to Grids
GdX - Grid eXplorer

• GdX meeting October 12-13, 2006 Orsay

2
In memory of Amelia De VivoJune 15, 1967 June
21, 2006
3
parXXL Project overview
Client machine comfortable fine grained
development environment
Distributed server fine grained simulations
on coarse grained architecture
4
parXXL Cell Programming model
Coarse grained parallelism
Fine grained parallelism
parXXL
5
parXXL Execution support
SSCRAP
6
parXXL Optimized cell nets
Basic cell net creations

• Basic user code
• Automatic load balancing
• Neighbor cells are not grouped

1 PE
1 PE
Main function
1 PE
Ex Cubic cell net
Optimized cell nets (parcellnet)

• Optimized user/parcellnet code
• Explicit load balancing
• Neighbor cells are grouped

7
Relaxation perfs. on GdX machine
1 Speed up fine grained simulations
using more processors
2 Extensibility process larger problem on
more processors with similar cost for each cell
iteration
Scales until 420 106 Cells on 310 PEs
8
parXXL Future GdX experiments

• 2006 objectives
• Running 109 cell simulations.
• Large scale pipeline algorithms (fluid dynamics)
• combining data and control parallelism.
• GdX cluster ? GdX Grid emulator ? Grid5000
• ? Use wrekavoc for heterogeneity emulation
• ? Use network emulation.

• 2007 objectives
• Experiments of complete simulations of physical
• systems (LMOPS/LPMIA).
• Experiments on Grid5000.

9
parXXL Conclusion Perspectives

• A framework to implement fine grained
  applications on coarse grained architectures
• scales up to 310 PEs (relaxation benchmark)
• Optimized runtimes for shm mainframes for
  clusters

Future

• Implement global IO functions Hybrid cell
  communications
• Extend the cellular network library
  (parcellnet)
• Experiments on Grid5000
• Application to optic hot plasma simulations
  (ITER project)

10
parXXL Scaling Fine Grained Algorithm Design
to Grids
Questions?