Study of Cluster Computing on a High Speed Network Environment

根据研究计划，在高速网络环境下，对新的集群计算平台进行了实现实验和性能评估。本文的主要研究成果如下：1）提出了一种基于混合模型（共享内存/消息传递）的混合DSE来实现集群计算。在性能评估中，开发了PVM和MPI所使用的消息传递机制，以支持高性能实现，并且没有显著的开销。2）为了实现高效的共享内存访问，实现了软件缓存，以支持缓存DSE实现，其中网络延迟可以隐藏。在高速网络环境下的实现结果证实，在远程访问共享内存，软件缓存隐藏了网络延迟。此外，从应用的角度来看，Cached DSE在速度上比现有的DSE有了很大的提高，从而证实了Cached DSE的有效性。3）在内存一致性模型方面，采用顺序一致性（SC）模型和发布一致性（RC）模型进行了实现和实验。此外，RC模型还包括早期发布一致性模型和延迟发布一致性模型，其区别在于一致性控制执行的时间。在这两种模型中，因为我们可以在无效协议和更新协议之间进行选择，所以总共实现了6个一致性协议。实验结果表明，该模型是有效的，具有较高的性能。

As mentioned in the research plan, implementation experiments and performance evaluation were conducted to develop the platform of new cluster computing on a high-speed network environment. The following are the major results of the research :1) A Hybrid DSE was developed to implement cluster computing based on hybrid model (shared memory/message passing). In the performance evaluation, a message passing mechanism, as utilized by PVM and MPI,was developed to support high-performance implementation and there was no remarkable overhead observed.2) To implement an efficient shared memory access, a software cache was implemented to provide support for the Cached DSE implementation where the network latency can be hidden. The implementation results obtained over a high-speed network environment confirmed that during the remote access to shared memory, the software cache hides the network latency. Furthermore, from the applications point of view, Cached DSE exhibits a large improvement with regards to speed as compared to the existing DSE,and therefore confirmed the effectivity of Cached DSE.3) As to the memory consistency model, implementation and experiments were conducted using Sequential Consistency (SC) model and Release Consistency (RC) Model. Furthermore, the RC model consists the Eager Release Consistency Model and Lazy Release Consistency Model whose difference is in the timing of coherence control execution. In either model, because we can choose between invalidate protocol and update protocol, a total of 6 coherence protocols were implemented. The experiment results verifies that the Lazy Release Consistency Model is efficient and exhibits high performance.