| YH-SUPE (YinHe-Simulation Utilities for Parallel Environment) is an infrastructure for developing and executing the programs of parallel simulation. Parallel simulation based on the high performance computer is often applied to get the final results that acquired only after parallel simulation is executed for many times with different scenarios and parameters. If it is users'responsibility to start every sample program and configure initialized parameters, the simulation progress will be hampered with more work burden or human errors. Therefore, it is significant to design a management tool of parallel simulation runtime configuration for YH-SUPE, which provides functions of configuring sample parameters for simulation, dispatching simulation objects, runtime controlling and automatic callbacking of simulation results.In order to resolve the problems about manually configuring large numbers of task parameters, low efficiency and reclaiming a great deal of results for the simulations run in YH-SUPE, this dissertation focuses on the research of management technique for parallel simulation and runtime configuration. The main works and innovations of the dissertation are as follows:1. High performance computer nowadays is often characterized with multi-CPU and multi-core architecture, accordingly, YH-SUPE supports three kinds of communication mechanisms of TCP,MPI and shared memory for parallel simulation. With more replications and different load of each CPU core, traditional scheduling strategies would result in low efficiency communication or excessive load of CPU cores. In order to resolve this problem, this paper proposes an resource-aware, dynamic scheduling technique, which determines whether simulation task can be loaded through sustainably watching the utilization of all the CPU cores for each computer node in real time and customized CPU utilization threshold. For high throughput of simulation task, this technique also can assign new simulation tasks to one or more nodes of idle CPU core, to supply shared memory mechanism for more processes.2. A large number of sample parameters need to be configured, because large-scale parallel simulation task includes many samples. Except for manual configuration of sample parameters, this dissertation also poses the technique of a batch of configuring sample parameters through leading in and exporting sample files. The aim of this technique is to design a standard template. The user can modify and enlarge sample parameters according to the template's format after exported a file which includes simulation task parameters. This technique can leave out manual processes because the sample parameters of this file can be led into simulation task only once.3. The result files of multi-sample in the computer need to be reclaimed after simulation .The dissertation proposes the technique about how to automatically collect and merge the results on the basis of filtration mechanism to resolve that problem. On the one hand, multiple result files of each sample can be collected in a specified directory; On the other hand, the multiple result files can be merged into a bigger data file. Filtration mechanism is used to discard useless files when files is collected and merged.In addition, a management tool for parallel simulation runtime configuration is also introduced, which is based on Django framework, Python program language and B/s mode. The tool has functions of user management, simulation task and sample parameters configuration, sample dispatch, runtime control and data result callback, and so on. These can reduce lots of manual operations and increase the simulation efficiency greatly. The tool also have pretty visualization interface which is good at enlargement and maintainability. Experimental results show that operations are simplified, functions are completed, efficiency is higher and the requirement for parallel simulation runtime configuration of YH-SUPE is met by this tool. |
PDF Full Text Request