Model message-size dependence of inter-node bandwidth in analytical scalability models

Develop enhanced analytical performance models for broadcast and shuffle in distributed multi-GPU SQL processing that treat the effective inter-node bandwidth parameter B_n as a function of message size, thereby improving prediction accuracy as the number of machines increases and per-message sizes decrease.

Background

The paper proposes simple analytical models for shuffle and broadcast that assume a constant effective inter-node bandwidth B_n. Microbenchmark results show that achievable throughput depends on message size, and model–measurement comparisons indicate growing discrepancies as the cluster size increases and per-message sizes shrink.

Accounting for message-size effects in B_n would yield more accurate model-based projections and enable better model-driven resource scaling and planning for future networks and larger clusters.