Distributed machine learning systems train models via iterative updates between parallel workers and the parameter server. To expedite the transmissions, in-network aggregation of updates along with the packet forwarding at those programmable switches decreases the network traffic over these bottleneck links. However, existing in-network aggregation schemas do not adequately prepare the most suitable switches for various worker distributions and fail to capture the dynamic network status. Based on the status derived from in-band network telemetry, we aim to select the best switches upon the optimization we formulate with the objective of minimum transmission latency. Although the problem is actually a non-linear integer program, by adopting delicate transformations, a substitute with totally unimodular constraints and separable convex objective is then solved to obtain the integral optimum. We implement our in-network aggregation protocol and reconstruct in-band network telemetry protocol upon real devices, i.e., Barefoot Wedge100BF and Dell servers. Then, we evaluate the performance of our proposed AGG algorithm and the results indicate that the completion of related coflows decreases 40% on average compared with other strategies, improving at least 30% performance, compared with the state-of-the-art.

Edge computing enabled Intelligent Road Network (EC-IRN) provides powerful and convenient computing services for vehicles and roadside sensing devices. The continuous emergence of transportation applications has caused a huge burden on roadside units (RSUs) equipped with edge servers in the Intelligent Road Network (IRN). Collaborative task scheduling among RSUs is an effective way to solve this problem. However, it is challenging to achieve collaborative scheduling among different RSUs in a completely decentralized environment. In this paper, we first model the interactions involved in task scheduling among distributed RSUs as a Markov game. Given that multi-agent deep reinforcement learning (MADRL) is a promising approach for the Markov game in decision optimization, we propose a collaborative task scheduling algorithm based on MADRL for EC-IRN, named CA-DTS, aiming to minimize the long-term average delay of tasks. To reduce the training costs caused by trial-and-error, CA-DTS specially designs a reward function and utilizes the distributed deployment and collective training architecture of counterfactual multi-agent policy gradient (COMA). To improve the stability of performance in large-scale environments, CA-DTS takes advantage of the action semantics network (ASN) to facilitate cooperation among multiple RSUs. The evaluation results of both the testbed and simulation demonstrate the effectiveness of our proposed algorithm. Compared with the baselines, CA-DTS can achieve convergence about 35% faster, and obtain average task delay that is lower by approximately 9.4%, 9.8%, and 6.7%, in different scenarios with varying numbers of RSUs, service types, and task arrival rates, respectively.