Princeton University
Network Servers
|
Network Systems Group Princeton University Network Servers |
This project focuses on techniques to improve network servers and networking software. We consider performance optimizations, improved behavior under heavy load, and more robustness to abuse. Much of this work is at the boundary between networking and operating systems.
We have investigated the latency properties of the popular Apache Web Server and have applied our latency-reduction techniques to it. We find that by eliminating lock convoys that induce head-of-line blocking, we are able to reduce latency in Apache by more than an order of magnitude. [USENIX 2006].
We have developed a novel server design that allows both for the construction of very simple servers, as well as policy filters that can be used with new and existing servers in order to provide better security, fairness, and quality-of-service than other approaches. [NSDI 2006].
We have investigated the origins of server-induced latency to understand how to improve latency optimization techniques under various loads. Despite latency profiles that suggest standard queuing delays, we find that most latency actually originates from negative interactions between the application and the locking and blocking mechanisms in the kernel [SIGMETRICS 2004].
We have implemented a performance monitoring tool, called DeBox, that obtains precise information about OS behavior at low cost. DeBox can be used to debug and tune network application performance on complex workloads. Using DeBox, we are able to diagnose a series of problematic interactions between a web server and the OS, resulting in latency reductions ranging from a factor of 4 to 47 [Usenix 2004].
We have devised a new approach to protecting network servers from denial of service attacks. The approach, which we refer to as defensive programming, involves systematically injecting sensors and actuators into the code: the sensors watch for resource abuse and the actuators takes appropriate action should abuse be detected [OSDI 2002].
We have developed a process model for network servers (including software-based routers), and investigated the issues involved in scheduling CPU cycles on such a system [SIGMETRICS 2001].
Vivek S. Pai
Larry Peterson
Tiger Qie
Yaoping Ruan