Bandwidth-aware selective data multicast method

Abstract

The invention discloses a bandwidth-aware selective data multicast method, which specifically comprises the following steps of: when receiving nodes are selected in a multicast task, preferentially selecting the receiving node which can obtain the most bandwidth under fairness distribution according to available bandwidth and load; after basic multicast rate values of all multicast are calculated, greedy check is performed in sequence, and if the multicast rate is still met after a receiving node is selected again, the receiving node is selected; and further greedy and sequentially checking whether a residual bandwidth exists between each multicast source and the corresponding selected receiving node or not, and if so, increasing the bandwidth to the multicast transmission rate. According to the invention, the multicast transmission efficiency is improved, the time required by the network to complete the synchronization of the controlled partial data is reduced, and a data multicast transmission control method with higher performance is provided for applications such as distributed machine learning model training and the like.

Description

technical field [0001] The invention belongs to the technical field of data transmission, and in particular relates to a bandwidth-aware selective data multicast method. Background technique [0002] In scenarios such as data-parallel-based distributed machine learning model training, the working nodes participating in the training need to periodically synchronize the new models or model updates obtained by their respective training to achieve collaboration. Aiming at such communication requirements, some researchers have proposed various implementation schemes at present. One of the schemes is to let each working node transmit the learned new model or the data after model decomposition to other nodes in a multicast manner. Assuming that there are a total of n working nodes participating in training in a distributed system, some recent studies have shown that many distributed machine learning training tasks have the characteristic of tolerating only some nodes participating...

AI Technical Summary

Problems solved by technology

However, considering that the available bandwidth of different working nodes is often different, randomly selecting p receiving nodes for each multicast transmission will have a certain probability of selecting a working node with a lower bandwidth, and the network optimization effect that can be brought is not stable , resulting in the time required for parameter synchronization is usually still very large, there is significant room for optimization