Large-scale parallel system simulation method based on single-node simulation

Abstract

The invention provides a large-scale parallel system simulation method based on single-node simulation, wherein the method comprises the following steps: step 1, firstly, a large-scale parallel program runs in an existing computing system or a cloud platform according to an expected process scale and problem scale, and running information is acquired in a program running process; step 2, after necessary information supporting operation of a single process is collected, a user specifies configuration of a target high-performance computer to be simulated; step 3, the user selects a part of processes to be simulated, specifies process-node distribution, inputs the processes and node information into configuration, and starts a parallel system simulation system in a local host machine cluster;and step 4, the user starts the parallel program in the parallel system simulation system, and in the execution process, the calculation process, process scheduling and process-to-resource competition of the processes executed in the same node are accurately simulated by the single-node whole system simulation instance.

Description

technical field [0001] The present invention relates to a simulation method and a simulator system structure of a large-scale parallel computing system, in particular to a method for efficiently simulating and executing part of a process in a large-scale parallel program when the target hardware does not have it, and a simulator system structure . Background technique [0002] System structure simulator plays a very important role in the process of computer system design and implementation. In the early stage of design, the system design scheme can be determined through coarse-grained functional simulation; in the process of system development, the microstructure design can be further improved through fine-grained clock-accurate simulation; at the same time, the system software of the target system can also be synchronized Development, to speed up system development; in the process of synchronous software and hardware research and development, the performance prediction and...

AI Technical Summary

Problems solved by technology

Second, although the trace-driven method has strong scalability, it can only predict some simple performance indicators of a parallel program on the target system, such as running time, because it does not actually execute parallel programs. The scope of use of this method

The method based on the execution drive is to run large-scale parallel programs in reality, but because the small-scale cluster runs at a full scale, and the maximum number of running processes on each node is limited, the number of required hosts will increase with the number of processes. increases in size, which limits the scalability of the system

In addition, as the number of simulated processes increases, when it increases to a certain scale, the reduction ratio of the system will be unacceptable, making the entire system unusable

Third, although the method of using traditional architecture simulators is very flexible, due to the scale of massively parallel systems and the high reduction ratio of traditional architecture simulators for accurate simulation, when the system scale reaches a certain level, the simulation The reduction ratio of the system will be unacceptable, making the system unusable

However, at this stage, the existing research results on simulation methods for massively parallel systems are deficient.

Images