We model World Wide Web (WWW) servers and clients running over an ATM network using the ABR service. The servers are modeled using a variant of the SPECweb96 [1] benchmark, while the clients are based on a model by Mah [2]. The traffic generated by this application is typically bursty, i.e., it has active and idle periods in transmission. Due to the fact that the underlying TCP congestion windows remain open until a timeout expires, these windows may be used to send data in a burst when the application becomes active again. This raises the possibility of large switch queues if the source rates are not controlled by ABR. We study this problem and show that ABR scales well with a large number of bursty TCP sources in the system.

Complete contribution in text w/o figures (23,074 bytes) | Adobe Acrobat (42,421 bytes)

Presentation slides in Adobe Acrobat format (34,593 bytes)