Constructive specification and synthesis of agents for custom and cross RPC.

Abstract

Remote Procedure Call (RPC) is perhaps the most popular paradigm used today to build distributed applications. Many RPC semantics have been designed and implemented in recent years to meet application-specific requirements, and more RPC systems will be built as new distributed applications emerge. As a consequence, the term RPC has grown to include a range of vastly different protocols above the transport layer. In this dissertation, we argue that the cost of implementing a new RPC system is higher than it should be. Another consequence of having many RPC semantics is that programs using different RPC protocols cannot be interconnected directly, and building a solution for each case in a large heterogeneous environment is prohibitively expensive. This dissertation proposes an agent synthesis scheme to reduce RPC development costs and to provide a general solution to the cross-RPC communication problem. The agent synthesis scheme uses a synthesizer to generate implementations of RPC agents from constructive high-level descriptions. Synthesis allows new RPC systems to be prototyped rapidly and facilitates cross-RPC communication. To validate this idea, we have fully designed and implemented our scheme to synthesize RPC agents and to perform custom and cross-RPC communication. Our agent synthesis scheme is realized through three subsystems: a universal RPC toolkit (URPC toolkit), a protocol construction language (Cicero), and a runtime support system for agent synthesis and management (Nestor). The URPC toolkit provides semantics-independent library functions, allowing programmers to prototype diversified new RPC systems with minimal coding effort. Cicero provides language support for describing complex protocol implementations, especially those requiring multiple-thread support. Nestor is the subsystem integrating the URPC toolkit, Cicero and other software utilities to perform agent synthesis and management for cross-RPC communication. An evaluation of our prototype implementation shows that our synthesis approach provides a more general solution than existing approaches, and with lower software development and maintenance costs, while delivering reasonable cross-RPC performance and excellent custom RPC performance.