Parallel calculation system, and method and program thereof

摘要

A parallel calculation system includes a plurality of functional nodes and a plurality of network switches, which serve as part of a communication path for communication between the functional nodes and function as communication channels. As for the functional nodes, a functional node group made up of part or all of a plurality of the functional nodes that the parallel calculation system includes is managed as one group. The network switches include a plurality of input/output ports, which are ports for inputting and outputting the communication request, and a switch, which outputs the communication request that the input/output port inputs to the input/output port. The network switches manage a communication bandwidth for each of the communication channels between the functional nodes and control the switch.

主权项

1. A parallel calculation system comprising:
a plurality of functional nodes having a mathematical operation function, an input/output function or a storage function, or a combination of the functions, for carrying out information processing; and a plurality of network switches, which serve as part of a communication path for communication between the functional nodes and function as communication channels, wherein the parallel calculation system executes a plurality of processes, the functional nodes include: a functional component having a mathematical operation function, an input/output function or a storage function, or a combination of the functions, for carrying out information processing; a functional node input/output port that is a port for inputting and outputting a communication request transferred between the functional nodes for interactive communication between the functional nodes; and a functional node group management unit that manages, as one group, a functional node group made up of part or all of a plurality of the functional nodes that the parallel calculation system includes, the functional node group being assigned to one of the plurality of processes, and the network switches include: a plurality of input/output ports that are ports for inputting and outputting the communication request; a switch that outputs the communication request that the input/output port inputs to the input/output port; and a bandwidth management unit that manages a communication bandwidth for each of the communication channels between the functional nodes belonging to the functional node group and controls the switch, wherein a local communication request, which is a communication request for carrying out interactive communication only between a plurality of the functional nodes that make up the functional node group, is used for communication between a plurality of the functional nodes that make up the functional node group; a global communication request, which carries out interactive communication between all the functional nodes that make up parallel calculation system, is used for communication on the communication path; and the functional node group management unit includes a functional node management information holding unit that manages information used in converting the local communication request into the global communication request or the global communication request into the local communication request, and a communication request conversion unit that uses the information managed by the functional node management information holding unit to convert the local communication request, which the functional component inputs or outputs, into the global communication request, which the functional node input/output port inputs or outputs, or the global communication request into the local communication request.