| Popis: |
This dissertation argues that admission control should be applied as early as possible within a system. To that end, this dissertation examines the benefits and trade-offs involved in applying admission control to a networked computer system at the level of the network interface hardware. Admission control has traditionally been applied in software, after significant resources have already been expended on processing a request. This design decision leads to systems whose algorithmic cost is a function of the load applied to the system, rather than the load admitted to the system. By performing admission control at the network interface, it is possible to develop systems whose algorithmic cost is a function of load admitted to the system, rather than load applied to the system. Such systems are able to deal with excessive applied loads without exhibiting performance degradation. This dissertation first examines existing admission control approaches, focussing on the cost of admission control within those systems. It then goes on to develop a model of system behaviour under overload, and the impact of admission control on that behaviour. A new class of admission control mechanisms which are able to perform load rejection using the network interface hardware are then described, along with a prototype implementation using commodity hardware. A prototype implementation in the FreeBSD operating system is evaluated for a variety of network protocols and performance is compared to the standard FreeBSD implementation. Performance and scalability under overload is significantly improved. |
