摘要
SCI(IEEE1596-1992)高速互连协议以其极低延迟的特点被应用于许多关键领域。SCI软件反射存储网是基于SCI的软件分布式共享存储系统,主要用于集群并行计算和实时系统中的数据共享。结点通过读写物理上分布、逻辑上唯一的共享存储空间共享数据。任何结点写入数据时,数据将以一定的逻辑拓扑被传送到所有结点的物理内存上。传送数据的逻辑拓扑直接影响网络的写延迟,因此提出一种动态、低延迟的最优树逻辑拓扑,并在此基础上设计和实现了延迟低、易编程的SCI软件反射存储通信库-RFM。实验证明,提出的最优树逻辑拓扑设计大大降低了网络的写延迟,提高了网络的通信性能。
The key advantage of SCI interconnect protocol is to provide ultra-low latency. SCI software reflective memory network is an implementation of SCI-based software distributed shared memory. It is being used in the internal communication of parallel computing and real-time systems. Any node writes the shared memory space, the system will transfer the data to the physical memory on all nodes to ＂backup＂ the data. The logical topology of transferring the data greatly affects the performance. This paper proposes a low-latency, dynamic tree topology, called ＂optimal tree＂. Based the ＂optimal tree＂, RFM, the improved SCI software reflective memory system is designed and implemented. The experiments prove the ＂optimal tree＂ topology improves the performance of the whole network by significantly deduces write-latency.
出处
《微计算机信息》
北大核心
2008年第33期78-80,共3页
Control & Automation
基金
"等离子滤波器及其可调性研究"
国家自然科学基金资助项目(60371022)
关键词
反射存储
逻辑拓扑
可扩展一致性接口
Reflective Memory
Logical Topology
Scalable Coherent Interface.
