ON PROBE STRATEGIES FOR DYNAMIC MULTIMEDIA SERVER SELECTION Lisa Amini a and Henning Schulzrinne b a IBM T.J. Watson Research Center, Hawthorne, New York b Columbia University, New York, New York ABSTRACT While there has been extensive research into wide-area load balancing for Web servers, dynamic server selection issues for rich media have largely been ignored. We argue that streaming media imposes new challenges that are not met by current proposals for collecting and disseminating server and network path metrics. We define a model for quantitatively evaluating the network overhead of competing Internet measurement strategies, propose a novel strategy to better meet requirements for rich media, and use topologies representative of the Internet to show our strategy reduces total and average per link probe overhead by as much as 98%. We discuss additional advantages to our approach, including the ability to represent multiple, potentially streaming media specific, metrics; the ability to effect policy- based selection; and protection against denial of service attacks. 1. INTRODUCTION Deploying replicas of Web sites at distributed locations to reduce client access latency and enhance reliability has become commonplace. Server selection refers to the process of directing an application-level request to an appropriate server. The goal is generally to direct clients to network-proximal servers. Streaming media, due its very large object sizes, write-once- read-many nature, and high-bandwidth, isochronous delivery requirements, is an example of a service where delivery from edge-of-network servers is of particular advantage. However, existing server selection work [1,2,3] focuses on selecting from amongst a relatively small number of mirrors hosting Web pages and images. Dynamic server selection for delivery of rich media poses significant new challenges. For example, passive server selection schemes [1] collect response times from previous transactions and use this data to direct clients to servers. Clients are directed to servers that may provide poor service in order to collect performance data. While this is a reasonable approach when transactions are frequent and short-lived, as with Web pages and DNS queries, streaming media connections are generally long-lived. Directing a client to a distant or overloaded multimedia server, simply for the purpose of data collection, is unacceptable. Active server selection schemes have also been proposed [2,3]. Under active schemes, instrumentation to periodically probe end-to-end paths is distributed throughout the network. Generally, two-way probing, in which a message is sent to a remote server to elicit a response for which the round trip time (RTT) can be measured, is used. Based on the probe’s RTT, an estimated delay, or metric, is assigned to potential client-server paths. The advantage is that metrics can be collected without penalizing end-user performance. The disadvantage is in sacrificing timeliness or accuracy for efficiency and scalability. That is, the overhead for E2E probing scales on O(T 2 ) where T is the number of nodes (or Tracers) probing the network. If the number of servers is small, the probe overhead is not an issue. However, due to the end-to-end bandwidth required to transmit audio and video, multimedia Content Delivery Networks (CDN’s) are distributing thousands of servers throughout the Internet [4]. In such scenarios, the number of Tracers and the probe frequency must be limited to the minimum required to accurately report network distance on some time scale. The authors of [5] propose an efficient E2E probing scheme that satisfies the requirements for supporting large numbers of client-server pairs. However, the timeliness of such information is anticipated to be on the order of days and therefore, not reflect transient properties. In this paper, we evaluate schemes to actively collect and disseminate server and network path metrics for effective multimedia server selection. We demonstrate that intelligent probe strategies can reduce total probe cost by 54% to 98%, and average cost by 70% to 98%. We summarize a number of additional advantages, including the ability to represent multiple metrics; effect policy-based server-selection; and protect against denial of service (DoS) attacks. 2. GOALS Our goal is to create a scalable, Internet-wide service to effectively measure and disseminate server and network path metrics. While our proposal offers benefits applicable to any wide-area network measurement service or resource location problem, we focus on directing a client to a network-proximal multimedia server. A key deployment issue for dynamic server selection systems is in how the mechanism to direct clients to an appropriate server can be transparently inserted in the request-response cycle. Today’s CDN’s typically use specialized agents within Web servers or proxies, or Domain Name Servers (DNS) to redirect a client request to an appropriate server [6]. In this paper, we focus solely on efficiently measuring and distributing metrics so any such agent can effectively select an appropriate server. We set the following goals for our service: Metrics: Current schemes focus on packet delay. For streaming media, metrics such as packet loss ratio, path length, or even a timely binary indication of whether the path is available and uncongested may provide better insight than RTT