Affiliations: Johns Hopkins University, Department of Computer Science, Baltimore, MD 21218, USA E‐mail: {baruch, yidu, bkhan}@cs.jhu.edu | Bell Laboratories, Lucent Technologies, 101 Crawfords Corner Rd, Room 4G‐627, Holmdel, NJ 07733‐3030, USA E‐mail: [email protected]
Note: [] Also with MIT Lab. for Computer Science. Supported by ARPA/Army contract DABT63‐93‐C‐0038, DARPA/Airforce contract F30602961029, NSF Award ID CCR9700157, and ARO contract #DAAH04‐95‐1‐0607.
Note: [] Supported by DARPA/Airforce contract F30602961029.
Note: [] Partial support from Kaman Sciences Corp. at the Center for Computational Sciences, Naval Research Laboratory, Washington, DC, USA.
Abstract: In the future, global networks will consist of a hierarchy of subnetworks called domains. For reasons of both scalability and security, domains will not reveal details of their internal structure to outside nodes. Instead, these domains will advertise only a summary, or aggregated view, of their internal structure, e.g., as proposed by the ATM PNNI standard. This work compares, by simulation, the performance of several different aggregation schemes in terms of network throughput (the fraction of attempted connections that are realized), and network control load (the average number of crankbacks per realized connection). The simulation emulate a connection oriented network with a PNNI‐like hierarchical source routing algorithm. Our main results are: (1) minimum spanning tree is a good aggregation scheme; (2) exponential link cost functions perform better than min‐hop routing; (3) our suggested logarithmic update scheme that determine when re‐aggregation should be computed can significantly reduce the computational overhead due to re‐aggregation with a negligible decrease in performance.
