1 / 7

Non-Collective Communicator Creation

Non-Collective Communicator Creation. Ticket #286 int MPI_Comm_create_group(MPI_Comm comm , MPI_Group group, int tag, MPI_Comm * newcomm ). Non-Collective Communicator Creation in MPI . Dinan , et al., Euro MPI ‘11. Non-Collective Communicator Creation.

Download Presentation

Non-Collective Communicator Creation

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Non-Collective Communicator Creation Ticket #286 intMPI_Comm_create_group(MPI_Commcomm, MPI_Group group, int tag, MPI_Comm *newcomm) Non-Collective Communicator Creation in MPI. Dinan, et al., Euro MPI ‘11.

  2. Non-Collective Communicator Creation • Currently: Collective operation • Non-Collective: Create communicatorcollectively only on new members • Avoid bulk synchronization • Load balancing: Reassign processes fromidle groups to active groups • Overhead reduction • Multi-level parallelism, small communicators • Recovery from failures • Not all ranks in parent can participate • Compatibility with Global Arrays • Past: collectives using MPI Send/Recv 

  3. Non-Collective Communicator Creation Algorithm 0 1 2 3 4 5 6 7 MPI_COMM_SELF (intracommunicator) MPI_Intercomm_create (intercommunicator) MPI_Intercomm_merge 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 ✓ 0 1 2 3 4 5 6 7

  4. Non-Collective Algorithm in Detail • R Translate group ranks toordered list of ranks onparent communicator Calculate my group ID Left group Right group

  5. Evaluation: Microbenchmark O(log2 g) O(logp)

  6. Case Study: Markov Chain Monte Carlo • Dynamical nucleation theoryMonte Carlo (DNTMC) • Part of NWChem • Markov chain Monte Carlo • Multiple levels of parallelism • Multi-node “Walker” groups • Walker: N random samples • Load imbalance • Sample computation (energy calculation) is irregular, can be rejected • Regroup idle processes into active group • Preliminary results: 30% decrease in total application execution time T L Winduset al 2008 J. Phys.: Conf. Ser.125 012017

  7. Santorini straw vote to proceed with a formal reading:Yes ( 16 ), No ( 0 ), Abstain ( 1 )

More Related