The effects of high bandwidth networks on wide-area distributed systems

John Handley Shaffer, University of Pennsylvania

Abstract

High-bandwidth transmission systems, with data rates of 1.2 Gb/s and more are now available. In wide area networks (WANs), the bandwidth$\sp*$delay product characterizes network behavior in the face of non-trivial delays, and raises the difficulty of effective congestion and flow control. Since the speed of light is inviolable, propagation delay limits the performance of distributed systems across a WAN. Propagation delay can be viewed as the minimal time for a bit to traverse the network. The thesis experimentally demonstrates that there is a significant gain from high bandwidth for many applications. Applications use data units (ADUs) comprised of sequences of bits. The sequence is not complete until the last bit arrives; this arrival time, after the propagation delay is accounted for, is entirely a function of bandwidth. In this thesis, I characterize the extent to which increases in bandwidth may be traded against propagation delay, under the assumption that application response time is the relevant metric. My experiments were done on a platform I implemented to directly compare local ethernet performance with a wide area (544 mile round-trip), 155Mb/s ATM network. I ported MNFS, a network shared memory implemented on Sun OS, to the IBM RS/6000 running the AIX operating system. This required the implementation of a new concurrency control primitive, the Write First Swap atomic operation, to manage asynchronous interactions between the page daemon and the network shared memory system. I investigated a range of delays using an experimental ATM switch output port controller, OPCv2, prototyped by Bell Communications Research; these delays were correlated with experimental data collected directly from the A scURORA Gigabit Testbed WAN link between Penn and Bellcore. The experiments show that ADUs used by real distributed applications, such as a video-on-demand server and a parallel heat equation solver, using a distributed shared memory system, are large enough that applications perform comparably between the WAN and Ethernet environments. The experiments are summarized with construction of a general purpose predictive model for distributed applications. This thesis is the first thorough experimental investigation of bandwidth$\sp*$delay product tradeoffs using real applications on an operating 155 Mb/s ATM network of such scale.

Subject Area

Computer science

Recommended Citation

Shaffer, John Handley, "The effects of high bandwidth networks on wide-area distributed systems" (1996). Dissertations available from ProQuest. AAI9628007.
https://repository.upenn.edu/dissertations/AAI9628007

Share

COinS