Research And Implement On Adaptive Communication Mechanism Of High-performance RTI

The high level architecture of HLA-based distributed simulation technology has been widely used in various fields of national defense and economic construction. As the simulation of the scale and complexity of simulation grows largely, some users attempt to migrate HLA applications to high performance computers in order to improve the efficiency of simulation system. High performance hardware platform typically provides a variety of different levels of communication environments (such as shared memory, IB (Infiniband), Ethernet, etc), however traditional RTIs (Run-time Infrastructure) only support TCP/IP communication protocol, it is difficult to make full use of high performance computers with the efficient hardware communication support. Therefore research on high performance Adaptive communication mechanism has great theoretical and practical significance for improving the performance and scalability of HLA simulation.In this thesis, we have conducted an in-depth research on adaptive communication mechanisms oriented to HPC-RTI, after we did comprehensive analysis of traditional RTI communication mechanisms. The main contributions and innovations are archived as follow:1) First, most of the traditional RTI communication frameworks only support communication base on TCP/IP, design structure is relatively simple, which bring the difficulty to use high-performance computer hardware efficient hierarchical communication support effectively. Therefore, we propose a low-level communication-oriented framework for HPC environments, the communication framework for high-performance computing platform-centric communications environment including shared memory/IB/Ethernet. In the framework lower application automatically select the best means of communication, to minimize communication delays as much as possible, so as to provide framework support for HPC-RTI efficient communication.2) Traditional RTI platform support few means of communication, it is hard to take full advantage of shared memory communication in single HPC nodes and IB connection between HPC nodes. For that issue, the thesis made a HPC-RTI shared memory/IB/Ethernet adaptive communication algorithm, the algorithm use communication network information to determine the best mean of communication between the two sides in order to select the communication way among shared memory, IB and Ethernet adaptively.3) As one of the high speed communication mechanisms on Muti-core processor, shared memory can provide highest communication efficiency for communication between two processes on one same HPC node. The common way to use shared memory is that many processes share one same memory block for inter-communications. As the number of the process grows, the conflicts come more often, which result in more system lock waits and lower the efficiency of shared memory communication. Therefore, the thesis proposes a RTI Federates'communication algorithm based on distributed shared memory, effectively reducing shared memory access violations between federates. As it provides parallel access to shared memory, the distributed shared memory algorithm improves efficiency of the shared memory communications. Experimental results show when there is one federation with four federates in it on one four-core HPC node, as one federate updates its attributes while the rest three reflect them. Compared to the usual, the distribute algorithm could reduce the attributes delay by less than 63%.Based on the above, we designed and implemented a communication subsystems ACS (Adaptive Communication System) based on the adaptive communication mechanism. Experimental results show, compared to traditional RTI only using TCP/IP as the communication way, the RTI based on ACS reduces updates attributes delay in one single HPC node by 62% or more, increases Federates'attribute throughput in one single HPC node to more than 2.7 times larger.