Evaluation of Collective Communication Algorithms Targetingin-Network Computation-Enabled High-Performance Networks

Abstract

This thesis explores the impact of in-network computing (INC) on the performance of collective communication operations in parallel computing systems utilizing a fat-tree network topology. By integrating computational tasks within network switches, INC offers a method to enhance communication efficiency in data-intensive environments. This study compares the performance of key collective operations such as barrier synchronization, broadcast, scatter, gather, and reduce under traditional and INC-enabled approaches, evaluating metrics such as timing cost, bandwidth usage, and synchronization quality. The results reveal that INC significantly reduces timing cost and bandwidth usage in collective operations, leading to improved scalability and synchronization consistency, especially in large-scale networks. This work underscores INC’s potential to optimize collective communication in parallel computing systems, providing a promising avenue for improving performance and scalability in high-performance computing infrastructures.