Áòòòööò Òò Äääððòò Óó Ååøöö Blockin¹áòòù Aeaeøûóöö Ìóôóðóóóó×

Azer Bestavros John Byers Khaled Harfoush best s.bu.edu byers s.bu.edu harfoush s.bu.edu Computer S ien e Department Boston University Boston, MA 02215 Te hni al Report BUCS-TR-2001-010 May 2001 Abstra t The development and deployment of distributed network-aware appli ations and servi es over the Internet require the ability to ompile and maintain a model of the underlying network resour es with respe t to (one or more) hara teristi properties of interest. To be manageable, su h models must be ompa t, and must enable a representation of properties along temporal, spatial, and measurement resolution dimensions. In this paper, we propose a general framework for the onstru tion of su h metri -indu ed models using end-to-end measurements. We instantiate our approa h using one su h property, pa ket loss rates, and present an analyti al framework for the hara terization of Internet loss topologies. From the perspe tive of a server the loss topology is a logi al tree rooted at the server with lients at its leaves, in whi h edges represent lossy paths between a pair of internal network nodes. We show how end-to-end uni ast pa ket probing te hniques ould be used to (1) infer a loss topology and (2) identify the loss rates of links in an existing loss topology. Corre t, eÆ ient inferen e of loss topology information enables new te hniques for aggregate ongestion ontrol, QoS admission ontrol, onne tion s heduling and mirror site sele tion. We report on simulation, implementation, and Internet deployment results that show the e e tiveness of our approa h and its robustness in terms of its a ura y and onvergen e over a wide range of network onditions.