Measurement-based Performance of Network Elements and Services

Tijani Chahed and Pierre Vincent
Telecommunication Networks and Services Dept.
Software and Networking Dept.
Institut National des Télécommunications

Objective

The aim of this work is to measure the performance of various network elements and services, notably local communications using an Ethernet switch and/or an ATM switch, distant communications traversing routers as well as terminal equipment, for UDP, TCP and HTTP flows.

Methodology

Our methodology relies on direct, active measurement of test traffic generated through TrafGen, a traffic generator application that we developed for these purposes. TrafGen generates UDP and TCP test traffic following a client/server architecture. The application input parameters are datagram size, transmission rate, ^Å. The receiving side of TrafGen calculates the reception rate, the sender side calculates emission rate, RTT.

Results

Several scenarios are investigated.

First, we measure the performance of a terminal equipment, in terms of capacity of treatment, rate and CPU. This is achieved by a local loop setting of TrafGen on a single station with the receiving port activated. Two operating systems, Windows NT and Linux, are compared in this context. Our results show the difference of performances of Tcp-Ip stacks.

Second, we set the transmitting and receiving sides of TrafGen on two distinct stations which offers a higher performance figures than the previous case as local resources are doubled in this case. We compare the performance for both the receiver side listening process activate and not. In the latter case, ICMP messages indicate the inaccessibility of the receiving port which degrades the performance. Both TCP and UDP traffic streams are used and evaluated. The results shows the bandwidth between the stations

Third, we measure the performance of a multi-point scenario where N stations communicate with a single PC as well as hierarchical multi-point where the setting is reproduced on several scales. Again, TCP and UDP traffics are used. The obtained results show the performances of the multipoint application

Fourth, we measure the performance and capacity of treatment of an Ethernet switch and a router, with a focus on the processor and memory. Two cases of Ethernet subnetworks are particularly studied : Ethernet subnetworks of the same rate and those of different rates. The switches themselves perform both 'on the fly switching and store and forward scheme. We present the performances of 10-100 and Giga Ethernet switches.

We conclude by showing how to extend the use of our tool for the evaluation of different network equipment: routers, NAT, firewall and distant communications between providers.