A PFC-enhanced flow control method for lossless long-distance data transmission in WAN

Artificial intelligence-generated content (AIGC) and intelligent computing applications have grown quickly in recent years, necessitating the inexpensive and effective transfer of large volumes of training data across wide area networks (WAN). Deterministic networking (DetNet) technology for wide-area remote direct memory access (RDMA) access has recently drawn interest in the context of intelligent computing center connectivity. In IP-routed data center networks, RDMA is implemented via the RoCEv2 (RDMA over converged Ethernet v2) protocol, which depends on priority-based flow control (PFC) to create a Lossless network. However, PFC can cause poor application performance due to issues such as network congestion caused by burst traffic, PFC deadlocks, and so on. To overcome these problems, we propose an improved PFC-based flow control method called Reduce Speed Priority-based Flow Control (RPFC). The content is the data sending rate of the upstream device predicted by the downstream device. When burst traffic occurs, the system calculates the difference between the predicted value and the bandwidth, encapsulates the difference into pause frames, and instructs the upstream device to adjust the rate to maintain network stability. The downstream device side’s estimate of the data transmission rate is crucial to the RPFC mechanism. Initially, we examined data transmission rate prediction using existing timing models. Following multiple tests on real-world data, we found that directly predicting the data sending rate was ineffective. Firstly, we investigated the prediction of data transmission rate using existing timing models, and by conducting extensive experiments on a real data set, we found that direct prediction of data sending rate is not effective. To realize RPFC, we propose a prediction algorithm that combines LSTM and Attention mechanisms for time window aggregation grading (TWAG-LSTMA).Experiments show that, by conducting a large number of experiments on a real dataset, our model obtains a good effect.