Operating system structure and communication frameworks.

摘要

The characteristic complexity of system software has generally made it inflexible and difficult to manage. Consequently, traditional monolithic systems are being redesigned and decomposed using object-oriented principles. One important approach is the micro-kernel structure which separates monolithic kernel-based systems into kernel-level and user-level portions. Two fundamental problems regarding operating system structure remain. The first is the choice of the objects into which systems should be decomposed. To date, most structuring has been relatively coarse-grained. The second is performance. Generally, performance has suffered with increased decomposition.; This dissertation describes a project which addresses the issues of object-based decomposition and performance in the context of system software for communication. A cornerstone of the project was the development of MKnet, a prototype TCP/IP implementation for the IBM Microkernel. The prototype, derived from 4.4 BSD UNIX, separates the protocol implementation from the remainder of the operating system, demonstrating effective decomposition. The internal structure of the MKnet server is determined by a set of frameworks also derived from the BSD code. A fundamental goal of the prototype's development was to increase the structure and improve the flexibility of the original BSD implementation while maintaining its performance.; MKnet demonstrates the feasibility of isolating communication support from the remainder of the operating system. The prototype performs reasonably well in the context of moderate bandwidth networks such as Ethernet. The current server-based implementation is however somewhat less efficient than the monolithic kernel-based approach.; Experience with MKnet supports the following recommendations. Although no essential characteristics of microkernel-based systems were identified which necessarily hinder performance, task-based protection, characteristically used on microkernels, was found to be inefficient. Alternate mechanisms, such as library-based protection, must therefore be developed. In general, significant redesign will be required to develop service implementations which derive significant flexibility and manageability from their structure. One specific technology which must be perfected before highly structured, library-based systems become commercially feasible is the use of shared memory as a cross-domain transport.