hub server

Abstract

An embodiment of the present invention provides a hub server. The hub server includes: a configuration module, several hub ports, a memory, and a communication bus. The hub port includes a first hub sub-port and a second hub sub-port; the first hub The subports are connected to the memory, the communication bus is connected to all the second hub subports, the configuration module is connected to the communication bus, and the communication bus is connected to the memory; the first hub subport is used to receive the first data and send the first data to the memory; the configuration module obtains the second data according to the configuration information and the first data, and distributes the second data to each second hub sub-port; the memory is used to store the first data and the second data, and the embodiment of the present invention uses the Each hub port is directly connected to the device responsible for computing inside the hub server. During the data synchronization and update process, it reduces the data congestion caused by the only link directly connected to the server and improves system performance.

Description

technical field [0001] The present invention relates to the technical field of communication, and more specifically, to a hub server. Background technique [0002] In a distributed computing cluster, the completion of many tasks requires multiple computing nodes to aggregate parameter information or share parameter information. Taking the distributed machine learning system as an example, multiple computing nodes need to communicate with the server. Usually, multiple hosts are connected to the same local area network through switches to form a typical distributed computing cluster system. Each computing node completes a round of After calculation, the updated parameter value is submitted to the server; the parameter server collects and fuses the parameter update results from all nodes, and then distributes the parameter update results to all computing nodes for the next round of iterative calculation. [0003] figure 1 It is a schematic structural diagram of a distributed ...

AI Technical Summary

Problems solved by technology

[0003] figure 1 It is a schematic structural diagram of a distributed cluster system in the prior art, such as figure 1 As shown, multiple computing nodes 103 in the distributed cluster system in the prior art are connected to the server 101 through the switch 102, so during the parameter update and synchronization process of the distributed system, the server 101 and the computing nodes 103 generate One-to-many and many-to-one traffic patterns, these two patterns make the link directly connected to the server 101 easy to induce congestion, resulting in packet queuing or even loss, thereby delaying the transmission of related business flows and affecting system performance

[0004] How to solve the problem that the link directly connected to the server is easy to induce congestion has become an urgent problem to be solved

Images