Traffic- and Multi-Tenancy-Aware In-Network Aggregation Placement for Distributed Machine Learning
Distributed machine learning is an effective method to alleviate the intensive computation costs of training; however, it suffers from network bottlenecks while collecting local results. The recent advent of programmable data planes has opened a new avenue, in-network aggregation, which executes gradient aggregations in the middle of the network, resolving network bottlenecks, and further accelerates distributed machine learning. However, due to resource-constrained features of current programmable data planes, deploying in-network aggregation functionalities throughout the network would impose an unacceptable burden, posing a need for sophisticated deployment. In this paper, a problem of deploying in-network aggregation functionalities is studied to minimize the total network traffic in multi-tenant distributed machine learning. We formulate the problem as an integer linear programming (ILP) problem and prove its NP-hardness. Since finding the optimal solution using the brute-force method is extremely complicated, we propose a traffic-aware in-network aggregation placement algorithm based on a two-stage many-to-one matching game (denoted TAPINA-MG). The simulation results demonstrate that TAPINA-MG shows near-optimal performance with low complexity, achieving up to 22.5%, 38.9%, and 96.0% reduction for network traffic, maximum link utilization, and communication time, respectively, compared to state of the art, and effectively handles dynamic situations with minimal migration delay and comparable traffic performance.