基于共享存储的高可伸缩嵌入式集群模型

利用对称多处理机(SMP)作结点可为嵌入式集群带来更高的计算性价比,但多个并行和存储层次也会带来存储一致性、可伸缩性、性能差异等问题.提出一种基于共享存储的嵌入式集群模型LESC.该模型通过高度综合实现"计算单元-互连一致性模块-系统"三级高可伸缩结构,获得功耗成本有效性.LESC完成分布式共享存储的基本功能,其目录缓存一致性和扩展的共享存储机制改善了传统存储层次,并利用"共享存储虚拟网络"提供模块级的高效通信,避免了网络硬件开销,同时支持MPI编程.经该模型的真实系统平台测试,模块内MPI通信性能是传统嵌入式集群的3倍以上,单元间通信性能可达单元内性能的86%以上,Linpack测试其扩展性能在最差情况下接近理想值的70%.

Shared Memory Based Embedded Cluster Model with High Scalability

Deploying SMP nodes in embedded clusters makes a better cost-performance ratio, while multiple parallel and memory levels bring problems on memory coherency, scalability, performance gap, etc. The paper proposes a shared memory (SM) based embedded cluster model called LESC. The model implements a highly integrated and scalable three-tiered architecture with computing unit, interconnect module in coherency (IMC) and system to gain power and cost efficiency. LESC completes fundamental DSM functions, improves the traditional memory hierarchy by its directory-based cache coherence and expanded SM mechanism, uses "Shared Memory Virtual Network" to offer an efficient communication without hardware cost and supports MPI programming. The tests of real system based on the model show that MPI communication performance in IMC is more than triple that of traditional embedded clusters. Besides, the inter-unit communication performance can reach more than 86% of the intra-unit performance. Its Linpack scalable performance approaches 70% of the ideal value in the worst case.