Model-based collectives

Shared library libcpm_coll.so implements different model-based algorithms of collectives:

• LinInterp-based collective operations
• Hockney-based collective operations
• LogGP-based collective operations
• PLogP-based collective operations
• LMO-based collective operations

Command-line arguments:

• verbose verbose mode (default: no)
• sgv scatterv/gatherv mode
  • 0 propagation (default)
  • 1 broadcast
  • 2 allover
• model S communication model: Hockney, PLogP, LMO (required for model-based collectives)
• file S model data file (required for model-based collectives)

In order to preserve the original MPI interface, all model-based implementations use the global variables that provide model parameters. The interface of the implementation of a collective communication operation Y based on the model X includes the following components:

• int X_Y(standard args) is the model-based collective operation itself (for example, Hockney_Scatter_bfs_binomial_min)
• X_model* X_model_instance is a global variable providing the model parameters (must be available at all processors in the MPI communicator)
• int X_initialize(MPI_Comm, X_model* model), int X_finalize(MPI_Comm) are functions responsible for allocation and deallocation of the model instance at all processors. The model argument encapsulates the model parameters obtained either from a file or the model builder.

All model-based algorithms are divided into two groups: model-specific and generic.

Model-specific collectives depend on certain communication performance models, using parameters specific for these models only. For example, LMO_Gather_split_flat directly uses the LMO model global variable and its threshold parameters to split the medium size messages and perform a series of linear gathers with small messages, in order to avoid escalations of the execution time on the clusters with the TCP/IP communication layer. A model-specific collective operation Y is implemented in the following way:

int X_Y(standard args) {
    if (condition with X_model_instance->param)
        return ...;
}

Generic collectives depend on the prediction of the execution time of some communication operation (communication primitive); they are parameterized by predictions, which can be provided by any model. The communication primitive can be either the collective operation itself or some other simple operation. See Generic model-based collectives.

Tools

Part of MPIBlib

• MPIBlib/tools/collective - performs a universal or operation-specific collective benchmark with given accuracy and efficiency
• MPIBlib/tools/collective_test - verifies implementations of collective communication operations

Tests

• Test for model-based collectives
• Parallel matrix multiplication