Configurable Algorithms for All-to-All Collectives

MPI_Alltoall is a commonly used collective that allows a fixed-size data block to be exchanged between every pair of processes. The function can be implemented through a logarithmic number of point-to-point communication rounds, where the exact number of rounds and total data exchanged among processes depend on the log base (radix). This paper presents a mathematical foundation for studying all communication patterns for the all-to-all collective by developing parameterized formulas for total communication rounds and data exchanged. The model is used to narrow down a radix, √P (P: process count), that effectively balances latency and bandwidth concerns, yielding optimal performance-as also confirmed via evaluation on the Theta and Polaris supercomputers at ANL. We also present a novel two-layer tunable radix algorithm to take advantage of the shared-memory parallelism offered by modern systems. The algorithm decouples communication rounds into two phases that can be individually optimized to take advantage of the shared memory and high-speed interconnect separately. Our approach demonstrates improvements of up to 3.8× on Theta and 4.2× on Polaris over the vendor-optimized MPICH-based implementation of MPI_Alltoall for fast Fourier transform application.