Chimaera: A high-bandwidth network interface supporting cooperative tasks.

Abstract

The recent history of operating systems development has focused on the time-sharing paradigm. Given the characteristics of early computer use, time-sharing based operating systems provided the techniques to insure the fair distribution of computer resources among a number of users. Hardware development has brought us to the point where single-user workstations have become the most common computing platform. As all of the programs executed on a single-user workstation are for the benefit of a single-user, these programs may be seen as more cooperative than competitive. A number of trends in current operating system design and development have focused upon providing improved resource allocation among cooperative programs since many of the time-sharing derived methods are no longer appropriate. This dissertation focuses upon this trend and presents a new resource management paradigm structured around cooperative tasks and fine-grain resource management. In this context, the dissertation also examines existing architectural and operating system structures that may provide support for cooperative tasks. Specifically it focuses upon the effectiveness of the cache in supporting the processing of network data by an application. We have found, through experimentation, that the monolithic BSD kernel had significantly better cache and TLB utilization than the Mach microkernel. These results suggest three general rules for network subsystem design: (a) efficient TLB utilization; (b) sensitivity to the effects of context switches; and (c) minimization of buffer access self-interference. Finally, a new architecture is presented for a high-bandwidth network interface--Chimaera. As the components of the architecture may be implemented either as hardware devices added to the hardware network interface, or as very-low level software components added to the lowest levels of the network protocol stack, trade-offs can be made between flexibility and the level of support provided for fine-grain resource allocation decisions. The dissertation concludes with the results of simulation, and implementation details of both hardware and software Chimaera components.