1 Measuring Quality of Service in a Differentiated Services Domain with Linux Priyadarsi Nanda, Andrew James Simmonds, and Samson Lee Abstract— We have developed an independent test-bed of Linux based soft routers on which experiments have been con- ducted to measure Quality of Service in a Differentiated Services domain. This paper presents the first stage of an ongoing research effort on resource management mechanisms for next-generation networks. Our current results show proof-of-concept scenarios demonstrating how Differentiated Services expedited forwarding per-hop behavior can be used to reduce data loss and jitter even in the presence of heavy background traffic. We propose an agent facilitating end-to-end services by implementation of a policy- based Bandwidth Broker that is both scalable and robust. Index Terms— Diffserv, QoS, PHB, Bandwidth Broker I. I NTRODUCTION T HE advanced networking features with support for Qual- ity of Service (QoS) in recent Linux kernels (2.2.9+) provide a flexible platform to experiment with resource man- agement in a Differentiated Service (Diffserv) domain [1], [2]. A preliminary test-bed was configured with simplified Expedited Forwarding (EF) Per-Hop Behavior (PHB) [3]. Performance figures were measured in various experiments and the results are presented in this paper. In future work this will be extended to support more PHBs and carry out our research to achieve end-to-end QoS for different applications, both current and future Internet applications. In doing so, one of our most important objectives is to develop and test the resource management issue as experienced in the current Internet standard. In this paper, we focused on providing hard end-to-end QoS guarantees and concentrated on the EF PHB. Given a network link with known bandwidth, traffic marked as EF is given higher priority over traditional Best Effort (BE) traffic. A configured EF reservation rate is assigned to the link such that EF traffic always achieves guaranteed throughput up to, but no more than the reserved rate. When BE traffic is forwarded through the same node, it is able to utilize any unused EF reserved bandwidth. This means that the EF class is bounded and isolated whereas the BE class is able to borrow some of the unused bandwidth of EF. In this paper we present our Manuscript received July 2002; Accepted September 9 2002; revised October 14 2002; Presented at the International Conference on Information Technology, held in Bhubaneswar, India, December 21-24 2002; Published in the proceedings of this conference (Tata McGraw-Hill 2003); Reformatted October 2003 P. Nanda and A. Simmonds are with the Department of Computer Sys- tems, Faculty of Information Technology, University of Technology, Sydney, Australia. (e-mail: [pnanda, simmonds]@it.uts.edu.au) S. Lee is with the Faculty of Engineering, University of Technology, Sydney, Australia (e-mail: samson.lee@uts.edu.au) results from a number of experiments conducted with the test- bed that we designed. Our main focus was to monitor various QoS parameters such as throughput, jitter and loss. Section II presents our simple test-bed scenarios and de- scribes how we can build different topologies. Section III describes implementation of our Core and Edge router and presents results that we obtained to measure throughput, jitter and loss. In section IV, we present a brief discussion on our proposal to implement a Bandwidth Broker (BB) in each Diffserv domain and achieve end-to-end resource management in a scalable way. Section V of the paper gives future direction and conclusion. II. DESCRIPTION OF THE TEST-BED In the experiments for the first stage of this project, the main aim was to configure the test-bed for a Diffserv domain with all necessary Diffserv elements functional and then look specifically into the resource utilization by two classes of traffic. We found the resolution of the default Linux system timer to be insufficient for our purposes, and hence we modified the existing kernel to reduce the effects of timer granularity. The default packet scheduler was also changed by modifying the script in the kernel such that timestamps are now generated in hardware instead of soft-generation for each cycle. Various packages were recompiled to support Diffserv functionality. A typical scenario involves traffic streams (cur- rently UDP), which are sent from a number of senders to a 10 Mbps 100 Mbps Ethernet 100 Mbps Ethernet Background Sender eth0 192.168.1.3/24 Foreground Sender eth0 192.168.1.2/24 Receiver eth0 192.168.3.2/24 eth0 192.168.1.1/24 eth1 192.168.102.2/24 eth0 192.168.3.1/24 eth1 192.168.102.1/24 Edge Router Core Router Fig. 1. Network Topology