With the increase sizes of training datasets and models, the bottleneck in distributed machine learning (DML) training has shifted from computation to communication. To address this bottleneck, we propose an all-optical switching network architecture for accelerating the communication phase of DML training. Experimental results validate packets with error-free and 385 ns server-to-server low-latency communication at traffic load of 0.9. Small-scale DML training experiment deployed in the proposed architecture shows that Resnet50, Resnet101, and Vgg19 can be accelerated by 1.16x to 1.48x compared to electrical switching network. The proposed architecture demonstrates a 58.9% enhancement in cost efficiency and a 60.9% improvement in power efficiency compared to the 3-tier fat-tree architecture.

The rapid development of artificial intelligence training traffic has fueled significant interest in optical switching network deployment in intelligent computing and data centers with high switching capacity and low power consumption. Considering flexibility and switching efficiency, the time-slotted optical packet switches are broadly deployed for flexible transmission. However, these implements are facing two key challenges in improving bandwidth utilization. Firstly, the time synchronization accuracy directly affects the time-slots alignment, where coarse alignment results in a longer protection gap between optical packets. Such extensions deteriorate the bandwidth utilization. Secondly, the precision of the time synchronization affects time-slot alignment accuracy variation. The alignment uncertainty lengthens the protection gap to guarantee the correct data transmission. To overcome limitations in accuracy and precision inherent in conventional Precise Time Protocol deployments, this paper proposed a precise and stable time synchronization scheme for optical time-slotted switching networks characterized by a physically simplified clock distribution and syntonization network, reusing the existing network connections. With the proposed synchronization method executed, the system achieves a significant accuracy at 1.48E-10 s of root-mean-square. Moreover, the time synchronization process achieves a stable performance with a variation at 3.01 E-12 s over a 10000 s observation window in the long-term examination. Furthermore, experimental results demonstrate the enhanced time-slot alignment performance with a 1.07E-7 s protection gap achieving zero packet loss in a 25 Gbps optical switching network.

With the rapidly developing artificial intelligence, metaverse, and 5G applications, the traffic in the Data Center Network exploded in the past decade. Optical switches are implemented to forward packets to reduce the impact of such traffic proliferation. However, the lack of optical processing capabilities challenges the controller and scheduler to switch copious traffic. To mitigate the packet loss, multi-task learning LSTM algorithm is deployed to predict the proportion of traffic flows to each server in DCNs. The prediction results achieve the MSE of 6.61E-3. Compared to the traditional LSTM neural network, the proposed scheme reduces the MSE by 20.4%, which can guide the bandwidth scheduling for high utilization in DCNs.