MPI (Moldflow Plastics Insight) information scheduling method based on reinforcement learning under multi-network environment

Abstract

The invention discloses an MPI (Moldflow Plastics Insight) information scheduling method based on reinforcement learning under a multi-network environment, aiming at overcoming the defect of low practical application performance of a high-performance parallel computer, caused by the traditional circulating scheduling method. The method comprises the steps of: initiating parameters in a process of starting an MPI system, creating Cm<2> Q tables according to a multiple Q table combined method for a computing environment matched with m networks; continuously receiving an MPI information sending request sent by application in a process of starting the MPI system, determining a current information segment, then obtaining a current environment state, scheduling the current information segment to an optimal network according to the state information of historical empirical values stored in the Q tables; and finally, computing an instant reward value obtained by the scheduling and updating Q values in the Q tables. By adopting the invention, the problems that communication loads are distributed unequally, can not adapt to the network state dynamic change and have poor adaptability on the computing environment can be solved, and the practical application performance of the high-performance parallel computer is improved.

Description

technical field [0001] The invention relates to a message scheduling method in an MPI system under a multi-set communication network environment. Background technique [0002] In the field of high-performance computing, the inter-process communication of parallel applications mostly adopts the method of message passing, and a standard message passing interface (Message Passing Interface, MPI) is provided for the design of parallel programs. The specific component that implements the standard message passing interface is called the MPI system. In the high-performance parallel computer system, the MPI system is responsible for the inter-process communication of the parallel application program. When the computing power of a single computing node is not high, the high-performance parallel computer only needs to configure a communication network for communication between computing nodes. At this time, there is no problem of scheduling MPI messages on multiple sets of networks. ...

AI Technical Summary

Problems solved by technology

[0027] The technical problem to be solved in the present invention is to address the defects of the round-robin scheduling method, and propose an MPI message scheduling method based on reinforcement learning, which acquires and saves the experience value of the scheduling process through self-learning, and dispatches messages according to the current status and status of each network. Historical experience value dynamically selects the sending network, solves the problems of unbalanced communication load distribution, inability to adapt to dynamic changes in network status, and poor adaptability to computing environments, and improves the actual application performance of high-performance parallel computers

Images