Collective communication method based on physical location awareness

Abstract

The invention provides a collective communication method based on physical location awareness. The method includes the steps of: sensing the location of an objective process; when the objective process and a current process are located in a same physical chip, determining whether a local operation core is idle through query; and under the condition that the local operation core is idle, initiating a memory direct access operation through the local operation core, and directly writing communication data in an objective process main memory on the same physical chip, so that communication in the same physical chip no longer needs to transfer through a network interface. According to the collective communication method based on physical location awareness, when large-scale collective communication is performed, through sensing of the location of the objective process, based on a router chip routing algorithm, and according to a port-conflict-free principle, routing conflict during communication is minimized.

Description

technical field [0001] The present invention relates to the field of computer technology, and in particular to a collective communication method based on physical location awareness. Background technique [0002] The fat tree network is an efficient and scalable network architecture that uses multi-layer routing chips for interconnection in a tree structure. The fat tree network is favored by the industry due to its low cost and high scalability. However, due to the need for multi-layer switching chips for networking, when user programs need intensive message communication, there are communication performance problems caused by port conflicts on the switching chips. The problem of descent is one of the important problems that cannot be avoided in the collective communication optimization method. In addition, the characteristic of the heterogeneous many-core platform is that the computing core array has powerful aggregated communication capabilities, which also provides the ...

AI Technical Summary

Problems solved by technology

Existing full-switching communication implementation methods pay too much attention to the point-to-point communication behavior itself for the two-way bandwidth capability of the network, while ignoring the problem of performance scalability in large-scale environments, resulting in the inability to fully exploit the communication potential of the underlying network and the CPU itself.

Images