Optimizing collectives for size #8
Description
The HPC literature includes many collective algorithms optimized for specific network topologies and message sizes. Based on the message size, MPI implementations might choose completely different communication algorithms to efficiently utilize bandwidth on large messages and hide the latency for small messages, e.g., with pipelining.
We want to evaluate how such algorithms can be ported to serverless communication.
- Research and evaluate improved broadcast.
- Research and evaluate improved reduction.
- Research and evaluate improved scan.