Shaping the distributed end-system for global distributed applications

摘要

Advances in optical networking technology provide the network with unique storage capabilities enabling the creation of a physical shared memory in a wide area distributed system. This has implications on the overall systems' architecture as well as on the host. From the user perspective the end-system will be shaped by the application and become a global distributed shared-memory multiprocessor. In this paper we propose a memory building block based on optical networking technology that we further use in devising a multiprocessor host architecture. A hierarchical memory architecture that ensures scalability and allows multiple copies of a shared object is adopted. The architecture favours hardware solutions for enforcing cache coherence by exploiting the intrinsic broadcast capability of the new memory module. The interprocess communication system suitable for such an environment is analysed in order to achieve functional transparency for applications ported from a multicomputer to a multiprocessor, and issues related to performance transparency when dealing with long light-propagation latency are addressed. With this processing-communication paradigm, the host-network interface simply provides a direct path between the host-memory and the external network-memory.