High performance distributed objects using caching proxies for large scale applications

Initial implementations of middleware based on standards such as CORBA have concentrated on host and language transparency issues in order to demonstrate interoperability. They have largely adopted a no-replication approach and have frequently neglected performance-at-scale issues. This has led to continuing deployment of either non-scalable full-replication approaches or ad-hoc messaging-based middleware for applications such as intelligent networks, WWW applications and collaborative virtual reality. These applications require millions of objects globally distributed across hundreds of hosts and demand a very high throughput of low-latency method invocations. Our main research aim is to be able to reason about the performance of such applications when using scalable partial-replication and object-oriented approaches to middleware. Our approach is to use a simulator to explore potential design and implemention choices. Our current simulator-driven design, called "MinORB", has been fully implemented and tested. MinORB supports scalable high performance by a combination of techniques, including weak and application-specified consistency and partial replication using fine-grained proxy caching. Experimental results show that our work compares very favourably with other leading implementations, such as OmniORB. Scalability is unparalleled, with up to 1,000,000,000 objects per address space, a maximum throughput of 42,000 invocations per second and service times as low as 4 ms.