Scalable startup issues in OpenMPI

1
Scalable startup issues in OpenMPI

• Camille Coti
• ICL retreat 2007

2
Problematic

• Startup time of an application
• Initialize the environment
• Spawning the daemons and distribute the
  application information
• Start the application itself
• Benchmarks dont measure initialization
• Scale
• Large
• Small
• What about medium?

3
First measurments

• Methodology time a trivial MPI application
• int main(int argc, char arg)
• MPI_Init(argc, argv)
• MPI_Finalize()
• Return EXIT_SUCCESS
• Increase the number of procs

4
Where are we losing time?

• Finer-grain timers
• In the application MPI_Finalize() is soooo quick
• MPI_Init() is the guilty guy
• In MPI_Init() time each stage

5
Whats happening in stage 1?

• Initialize the system
• Exchange some information with the GPR ?
  bottleneck!
• Each daemon initializes its network interfaces
  and internals
• Then it sends its contact information to the
  registry/seed/GPR
• GPR broadcasts those information to everybody
• Then they can contact each other ? start stage 2

6
How to improve information exchange

• Try different topologies
• Direct
• Linear
• Binomial tree
• Ring (new!)
• Overlap communication / processing

7
Future Works

• Work with Jelena and Thara
• Merge the BMG with the collective communications
  algorithms and the ORTE knowledge the magic
  potion.
• Reuse the already existing knowledge dont
  reinvent the wheel
• Soon will be the time to go back to Paris
• Eat plenty of foie gras
• Continue the remote collaboration
• See you back in March 2008