Distributed computing system, node device, and optical network switching method

Abstract

Efficiently utilizing multiple optical fiber networks with different topologies. [Solution] A distributed computing system is provided that includes multiple nodes and multiple optical network lines. Each node includes multiple optical switches, multiple optical transistors, a data processing unit, and a network control unit. Each optical switch has a first port connected to an optical transistor and multiple second ports to which multiple optical paths connected to multiple optical network lines are connected, and the second port to which the first port is connected can be switched. The network control unit of each node simultaneously switches the connection state of the multiple optical switches when the aggregate communication currently being performed is completed.

Claims

1. A distributed computing system comprising multiple nodes and multiple optical network lines, wherein the multiple optical network lines have different network topologies. Each of the plurality of nodes includes an optical switch array containing a plurality of optical switches, a plurality of optical transceivers corresponding to each of the plurality of optical switches, a processing unit for processing data input and output via the plurality of optical transceivers, and a network control unit for controlling the switching of the plurality of optical switches. Each of the aforementioned optical switches has a first port to which a corresponding optical transceiver is connected, and a plurality of second ports to which a plurality of optical paths connected to the plurality of optical line networks are connected, and the second port to which the first port is connected can be switched. In each of the plurality of nodes, when the aggregate communication being performed by the node in the current connection state of the plurality of optical switches is completed, the network control unit simultaneously switches the connection state of the plurality of optical switches. Distributed computing system.

2. Each of the plurality of nodes further includes a plurality of light intensity tuners, each of which is positioned on the plurality of light paths, and each of the plurality of light intensity tuners has a first mode for allowing the light signal to pass through and a second mode for suppressing the passage of the light signal. In each node, the network control unit shall The optical intensity adjuster on the optical path used for the collective communication is operated in the first mode, and the optical intensity adjuster on the optical path not used for the collective communication is operated in the second mode. The distributed computing system according to claim 1.

3. At least one of the aforementioned plurality of optical fiber networks is a topology-variable optical fiber network capable of changing the network topology. A distributed computing system according to claim 1 or 2.

4. The multiple network topologies corresponding to the multiple optical fiber network include a ring network topology and a fully connected network topology in which every two of the multiple nodes are connected to each other. The optical network corresponding to the ring-type network topology is used for data-parallel AI (Artificial Intelligence) model learning processing, and the optical network corresponding to the fully connected network topology is used for tensor-parallel or expert-parallel AI model learning processing. A distributed computing system according to claim 1 or 2.

5. In a distributed computing system including multiple nodes and multiple optical network lines, a node device functions as one of the multiple nodes, wherein the multiple optical network lines have different network topologies. The node device includes an optical switch array containing a plurality of optical switches, a plurality of optical transceivers corresponding to each of the plurality of optical switches, a processing unit that processes data input and output via the plurality of optical transceivers, and a network control unit that controls the switching of the plurality of optical switches. Each of the aforementioned optical switches has a first port to which a corresponding optical transceiver is connected, and a plurality of second ports to which a plurality of optical paths connected to the plurality of optical line networks are connected, and the second port to which the first port is connected can be switched. When the network control unit completes collective communication performed by the node device in the current connection state of the plurality of optical switches, it simultaneously switches the connection state of the plurality of optical switches. Node device.

6. A method for switching optical network connections in a distributed computing system including multiple nodes and multiple optical network connections, wherein a node device functioning as one of the multiple nodes has different network topologies, The node device includes an optical switch array including a plurality of optical switches, a plurality of optical transceivers corresponding to each of the plurality of optical switches, a processing unit for processing data input and output via the plurality of optical transceivers, and a network control unit for controlling the switching of the plurality of optical switches, wherein each of the plurality of optical switches has a first port to which a corresponding optical transceiver is connected, and a plurality of second ports to which a plurality of optical paths connected to the plurality of optical line networks are connected, and the second port to which the first port is connected is switchable. The method includes the step of simultaneously switching the connection states of the plurality of optical switches when the network control unit completes collective communication performed by the node device in the current connection states of the plurality of optical switches. How to switch to a fiber optic network.

Images