Performance Properties of Combined Heterogeneous Networks Nader Mohamed, Jameela Al-Jaroodi, Hong Jiang, and David Swanson Department of Computer Science and Engineering University of Nebraska – Lincoln Lincoln, NE 68588-0115, [nmohamed, jaljaroo, jiang, dswanson@cse.unl.edu] Abstract In this paper we study the performance properties of combined multiple heterogeneous networks. Consider two nodes node1 and node2 with multiple heterogeneous physical networks connecting them. Each network has some properties such as bandwidth, latency, and packet loss ratio. We develop an analytical model to find the peak performance, or upper bound, of combined networks for the transfer of a message of a given size in parallel. The model is based on the existence of an optimal load balancing mechanism for parallel transfer that fully utilizes the existing networks. The paper provides a framework to find the peak performance parameters of message transfer over combined heterogonous networks as well as to measure the actual utilization and efficiency of parallel message transfer implementations. 1. Introduction Fully utilizing the collective bandwidth of multiple networks existing between two or more machines to perform parallel transfer of large files and messages has been an active research area [1, 2, 3, 4, 5, 10 and 11]. Parallel transfer over multiple networks can reduce total transfer time as well as increase the effective bandwidth. Applications such as remote satellite observation, distributed data mining, and distributed scientific simulation experiments, such as high-energy physics simulation experiments [9], are both compute intensive, requiring scalable high-processing power, and data intensive, demanding reliable, scalable high-bandwidth communication infrastructure for high-volume and high- speed data access. With the advancement of technology, computational grids [7] are poised to provide the most suitable infrastructure for these distributed applications. One of the requirements of these applications is high-speed transfer and data replication to enhance both data access and locality. One approach to increase transfer speed is by adding more network connections such that they increase the total communication bandwidth among the grid components. Grid resources are geographically distributed and may be connected by multiple homogeneous or heterogeneous networks. Each network has some defining properties such as bandwidth, latency, and packet loss ratio. The bandwidth is measured by the number of bytes transferred end-to-end every second. The latency represents the total amount of time it takes a single byte to travel from the source to the destination, including all intermediate delays. The packet loss ratio is the percentage of packets lost during transmission. For example (in Figure 1), there are two heterogeneous networks between node1 and node2 in a Grid. The bandwidth, latency, and packet loss ratio for the first network are B 1 , l 1 , and p 1 , respectively, and for the second network they are B 2 , l 2 , and p 2 , respectively. There are some efforts to develop network services that logically combine heterogeneous networks to appear as a single network for distributed applications. One example is MuniSocket [10], which will be briefly discussed in section 2. In order to evaluate and compare the actual performance of the different protocols and implementations that utilize combined heterogeneous networks, the peak performance of combined networks should be known. In this paper, without assuming any specific implementation or protocol of parallel transfer over multiple heterogeneous networks, we develop a mathematical model to express the theoretical performance parameters of combined networks. These networks may be two dedicated networks or two disjoint paths in a wide area network (WAN) that connect both nodes. In our research, we consider end-to-end connectivity and thus the intermediate components such as routers and switches in the network path (if present) are not directly addressed. However, their effects are considered as part of the latency measured. To derive the theoretical performance properties of combined heterogeneous network we assume the existence of an optimal load balancing parallel transfer technique that is aware of the properties of the networks 0-7695-1926-1/03/$17.00 (C) 2003 IEEE Proceedings of the International Parallel and Distributed Processing Symposium (IPDPS’03)