Check out the new USENIX Web site. next up previous
Next: Conclusions and Future Work Up: Kernel Mechanisms for Service Previous: Comparison of User Space


Related Work

Several research efforts have focused on admission control and service differentiation in web servers [19], [20], [21], [22], [8] and [23]. Almeida {\em et al.} [8] use priority-based schemes to provide differentiated levels of service to clients depending on the web pages accessed. While in their approach the application, i.e., the web server, determines request priorities, our mechanisms reside in the kernel and can be applied without context-switching to user level. {\em WebQoS} [23] is a middleware layer that provides service differentiation and admission control. Since it is deployed in user space, it is less efficient compared to kernel-based mechanisms. While WebQoS also provides URL-based classification, the authors do not present any experiments or performance considerations. Cherkasova {\em et al.} [20] present an enhanced web server that provides session-based admission control to ensure that longer sessions are completed. Crovella {\em et al.} [24] show that client response time improves when web servers serving static files serve shorter connections before handling longer connections. Our mechanisms are general and can easily realize such a policy.

Reumann {\em et al.} [25] have presented virtual services, a new operating system abstraction that provides resource partitioning and management. Virtual services can enhance our scheme by, for example, dynamically controlling the number of processes a web server is allowed to fork. In [26] Reumann {\em et al.} have described an adaptive mechanism to setup rate controls for overload protection. The receiver livelock study [2] showed that network interrupt handling could cause server livelocks and should be taken into consideration when designing process scheduling mechanisms. Banga and Druschel's [27] {\em resource containers} enable the operating system to account for and control the consumption of resources. To shield preferred clients from malicious or greedy clients one can assign them to different containers. In the same paper they also describe a multi listen socket approach for priorities in which a filter splits a single listen queue into multiple queues from which connections are accepted separately and accounted to different principals. Our approach is similar, however, connections are accepted from the same single listen queue but inserted in the queue based on priority. Kanodia {\em et al.} [21] present a simulation study of queuing-based algorithms for admission control and service differentiation at the front-end. They focus on guaranteeing latency bounds to classes by controlling the admission rate per class. Aron {\em et al.} [28] describe a scalable request distribution architecture for clusters and also present resource management techniques for clusters.

Scout [29], Rialto [30] and Nemesis[31] are operating systems that track per-application resource consumption and restrict the resources granted to each application. These operating systems can thus provide isolation between applications as well as service differentiation between clients. However, there is a significant amount of work involved to port applications to these specialized operating systems. Our focus, however, was not to build a new operating system or networking architecture but to introduce simple controls in the existing architecture of commercial operating systems that could be just as effective.


next up previous
Next: Conclusions and Future Work Up: Kernel Mechanisms for Service Previous: Comparison of User Space
Renu Tewari
2001-05-01