Master selection method based on multiple computing devices

Abstract

The invention relates to a master selection method based on multiple computing devices, which comprises the following steps: S1, when a server is powered on, acquiring topological structure information {G1, G2,... GN} of the computing devices, if G1, G2,... GN are sequentially connected to form a linear topological structure, executing the step S2, generating a request data packet Ri by each Gi, the Ri comprising id of the Gi, and sequentially transmitting the Ri to G1 from Gi-1 along a first direction of the linear topological structure by the Gi, the Gi sequentially transmits the Ri to the GN from the Gi + 1 along the second direction of the linear topological structure; after the Gj receives the Ri, a reply data packet Qij is generated, the Qij comprises the id of the Gj and a target parameter of the Gj, and the Gj transmits the Qij to the Gi along a linear structure; and S3, obtaining the number Mi1 of the Qij received by the Ri sent along the first direction of the linear topological structure and the number Mi2 of the Qij received by the Ri sent along the second direction of the linear topological structure, and if Mi1 = Mi2, or Mi1 + 1 = Mi2, determining the Gi as master. According to the invention, the information interaction efficiency among a plurality of computing devices is improved.

Description

technical field [0001] The invention relates to the technical field of computers, and in particular to a method for selecting a master based on multiple computing devices. Background technique [0002] AI computing devices such as GPUs, FPGAs, DSPs, and AI accelerators in servers need to maintain power consumption near Thermal Design Power (TDP) during normal operation. Existing servers or server clusters usually include For multiple computing devices, how to reasonably allocate power consumption to the multiple computing devices is very important during the operation of the multiple computing devices. In the prior art, a master (main computing device) is usually selected among multiple computing devices, and power consumption is allocated to the multiple computing devices through the master. However, there are at least the following shortcomings in the prior art: (1) The prior art usually randomly selects the master, which may lead to a long distance between some computing...

AI Technical Summary

Problems solved by technology

However, there are at least the following disadvantages in the existing technology: (1) The existing technology usually selects the master at random, which may cause some computing devices to be far away from the master, resulting in a large delay, which affects the efficiency of power allocation. Considering the delay of each computing device, however, since the number of computing devices may change, the original master can no longer guarantee a reasonable delay after the number of computing devices changes, which in turn will affect the power consumption between multiple computing devices distribute

(2) In the prior art, in the process of allocating power consumption for multiple computing devices, the static allocation is usually only based on the TDP value of the computing devices, without fully considering the real-time working status of each computing device, thus resulting in waste of power consumption , low power utilization

Images