Scalability in Distributed Systems, Parallel Systems and Supercomputers

摘要

With the dramatic improvement in communications, the trend is to move from specialised machines (supercomputers) that are optimised for one specific application and used by a single user [16,20] to general-purpose, high-performance machines that perform reasonably well on a variety of applications [17,21]. These machines are based on powerful processors interconnected by a high bandwidth interconnection (e.g. Fiber-Channel, Hippi) which are available. As a result our view of distributed systems, parallel systems and even supercomputers converges and it becomes possible to look at some issues in a unified manner. In this paper scalability is addressed and design principles are suggested to meet the scalability goal. Usage of these principles is demonstrated with a server selection (trading) service. Adaptive partitioning is advocated as a key design principle for scalability and is compared to the same service provided using other partitioning techniques.