Session 6 Track A: Challenges and Extensions
- Full paper 7: A Library for Advanced Datatype Programming
Jesper Larsson Träff - Full paper 8: On the Expected and Observed Communication Performance with MPI Derived Datatypes
Alexandra Carpen-Amarie, Sascha Hunold and Jesper Larsson Träff - Full paper 9: MPI Sessions: Leveraging Runtime Infrastructure to Increase Scalability of Applications at Exascale
Daniel Holmes, Kathryn Mohror, Ryan E Grant, Anthony Skjellum, Martin Schulz, Wesley Bland and Jeffery M Squyres
Session 6 Track B: Parallel Applications using MPI
- Full paper 10: Distributed Memory Implementation Strategies for the kinetic Monte Carlo Algorithm
Antonio Esteves and Alfredo Moura - Full paper 11: How I Learned to Stop Worrying and Love In Situ Analytics: Leveraging latent synchronization in MPI collective algorithms
Scott Levy, Kurt Ferreira, Patrick Widener, Patrick Bridges and Oscar H. Mondragon - Short paper 1: The Potential of Diffusive Load Balancing at Large Scale
Matthias Lieber, Kerstin Gößner and Wolfgang E. Nagel
Full paper 7: A Library for Advanced Datatype Programming
Jesper Larsson Träff
We present a library with functionality beyond the MPI standard for manipulating data layouts described by derived datatypes. Main contributions are: a) Constructors for several, new datatypes for application relevant layouts. b) New navigation and query functionality. c) Representation of type signatures by explicit signature types. As a simple application, we implement reduction collectives on non-contiguous data. The functionality could be implemented more efficiently within an MPI library.
TOP_______________________________________________
Full paper 8: On the Expected and Observed Communication Performance with MPI Derived Datatypes
Alexandra Carpen-Amarie, Sascha Hunold and Jesper Larsson Träff
We examine natural expectations on communication performance using MPI derived datatypes, in comparison to baseline, ``raw'' performance for communication of simple, non-contiguous data layouts. We show that common MPI libraries violate these datatype performance guidelines, and discuss reasons why this happens, but also show cases where MPI libraries perform well. Our findings are in many ways surprising and disappointing.
TOP_______________________________________________
Full paper 9: MPI Sessions: Leveraging Runtime Infrastructure to Increase Scalability of Applications at Exascale
Daniel Holmes, Kathryn Mohror, Ryan E Grant, Anthony Skjellum, Martin Schulz, Wesley Bland and Jeffery M Squyres
MPI includes all processes in MPI_COMM_WORLD; this is untenable for reasons of scale, resiliency, and overhead. This paper offers a new approach, extending MPI with Sessions, which makes two key contributions: a tighter integration with the underlying runtime system; and a scalable route to communication groups. This is a fundamental change in how we organise and address MPI processes that removes well-known scalability barriers by no longer requiring the global communicator MPI_COMM_WORLD.
TOP_______________________________________________
Full paper 10: Distributed Memory Implementation Strategies for the kinetic Monte Carlo Algorithm
Antonio Esteves and Alfredo Moura
The paper presents distributed memory strategies to parallelize a kinetic Monte Carlo algorithm, exploring different computation vs. communication ratios. The simulated process is the precipitation in an Al-Sc alloy. Results prove that our parallel implementation reproduces the statistics of the sequential kMC. The simulation time was reduced with a distributed parallelization but, the non-deterministic nature of kMC obliges to introduce approximations to achieve significant and scalable gains.
TOP_______________________________________________
Full paper 11: How I Learned to Stop Worrying and Love In Situ Analytics: Leveraging latent synchronization in MPI collective algorithms
Scott Levy, Kurt Ferreira, Patrick Widener, Patrick Bridges and Oscar H. Mondragon
To mitigate the cost moving large volumes of data from compute nodes, many applications rely on in situ data analytics. Processing output data on the nodes where it is generated reduces pressure on the I/O subsystem. In this paper, we show that leveraging synchronization induced by some implementations of MPI collective operations to schedule analytics codes may significantly the reduce the impact on application performance.
TOP_______________________________________________
Short paper 1: The Potential of Diffusive Load Balancing at Large Scale
Matthias Lieber, Kerstin Gößner and Wolfgang E. Nagel
Dynamic load balancing with diffusive methods is known to provide minimal load transfer and requires communication between neighbor nodes only. These are very attractive properties for highly parallel systems. We compare diffusive methods with state-of-the-art geometrical and graph-based partitioning methods on thousands of nodes. When load balancing overheads, i.e. repartitioning computation time and migration, have to be minimized, diffusive methods provide substantial benefits.
TOP_______________________________________________
