Fast and Scalable Barrier Using RDMA and Multicast Mechanisms for InfiniBand-Based Clusters

Approximating the buffer allocation problem using epochs

“…Consequently, we briefly discuss their implementation here. In MPI the MPI_Barrier is implemented using the Butterfly or Pair Exchange (PE) [18,24] algorithm. This can easily be visualized by letting each of the n processes be represented by a vertex in a hypercube of dimension log n, The barrier is performed in log n stages: in the first stage each process synchronizes with their neighbour along the first dimension of the hypercube, in the second stage each process synchronizes with their neighbour along the second dimension, and so on.…”

A practically constant-time MPI Broadcast Algorithm for large-scale InfiniBand Clusters with Multicast

“…The multicast operation has also been applied to implement other collective operations like MPI BARRIER, MPI ALLREDUCE or MPI SCATTER [2,7,9]. We use the scheme proposed in [19] for MPI BCAST and adapt it for the use with the InfiniBand TM multicast technology.…”

Energy, Memory, and Runtime Tradeoffs for Implementing Collective Communication Operations

“…chip, Panda et al demonstrate a series of algorithms tuned to InfiniBand T. Hoefler, D. Moor networks and RDMA systems [27,42], and Almasi et al [3] show optimization techniques for the BlueGene/L Torus network.…”