Hyper MPI
Based on Open MPI and OpenUCX, Hyper MPI supports MPI-3.1 and Unified Communication Group (UCG). It boosts communication efficiency by using optimized collective communication algorithms, high-speed inter-node interconnects, intra-node shared memory, and hardware offloading.
Allreduce
- Performs element-wise arithmetic (such as sum and product) or logical operations (like AND or OR) across send buffers from all processes, then synchronizes the result to the receive buffers of every process in the communicator.
Bcast
- Broadcasts data from a root process to all other processes within the communicator, ensuring all processes receive the same data.
Alltoallv
- Performs a point-to-point data exchange where every process sends variable-length data blocks to every other process, and receives data blocks in return.
Allgatherv
- Collects variable-length data blocks from all processes and distributes the combined result to every process in the communicator.
Gatherv
- Controls the root process to distribute data blocks to various processes in the communicator.
Scatterv
- Collects variable-length data blocks from all processes in the communicator.