• Project Description
• Project Activities
• Project Members
• Publications
  

As the client population accessing Internet services grows in size and dispersion, it is necessary to improve performance and scalability by deploying multiple, distributed server sites, each perhaps with its own cluster of server machines. Distributing services has the benefit of moving them closer (topologically) to end-users, thus reducing access latency, and improving service scalability by distributing the load among several sites.

One important issue in such a scenario is how to direct clients to appropriate service location. Choosing a server (or server site) arbitrarily may result in long delays due to distance or load on the server. Thus there must be some intelligent mechanism to select among a set of servers that can all deliver the requested service. Some desirable features of the server selection function include:

• client transparency -- avoid having the client know that there are choices, or at least automate the selection process
• service independence -- while WWW is the most pervasive server-based application, there are other important ones including distributed games, FTP, and Email.
• low overhead -- avoid introducing substantial additional network traffic or latency in the selection process
• support for multiple policies -- the server selection policy should be flexible and configurable
• wide-area load balancing -- users should be directed to servers in such a way that the server site load is roughly balanced; this is really just a special server selection policy but worth hilighting separately.
• high availability -- the server selection function should not be subject to single points of failure.

We are interested in exploring, architecting, prototyping, and evaluating general mechanisms for server selection and load-balancing that meet some or all of these characteristics.

DNS-based Server Selection

In focusing on the first feature above, i.e., client transparency, we are initially exploring DNS-based server selection mechanisms. The transparent nature of name resolution can be exploited to redirect clients to a an appropriate server without requiring any modification to client software, server protocols, or Web applications. Several commercial services (Akamai, Digital Island) and products (IBM, Cisco, F5, Alteon , Foundry) use DNS-based techniques for wide-area load-balancing and distributed server selection.

We have completed an initial analysis of two pertinent issues in DNS-based server selection. The first is the effect of reduced DNS caching on end-user latency. To achieve responsiveness to network and server load changes, DNS-based schemes must us small TTL values on resource records, thus requiring clients to perform additional name lookups. The second issue is the assumption that client and their local nameservers are proximal to each other, such that the nameserver is a good representative of the client location. We document our findings in the research report below.

We are currently building a prototype of a DNS-based server selection architecture which relies on passive nework performance monitoring to make decisions. This work is ongoing.

SPAND - Shared Passive Network Performance Discovery

We are modifying SPAND, a toolkit developed in collaboration with researchers at UC Berkeley. SPAND facilitates collection and sharing of network performance information by collecting performance reports in a local SPAND performance server. Others wishing to learn the performance experienced may then query the server for performance reports.

For more information, including software and publications, please see the SPAND homepage.

Anees Shaikh
Lisa Amini
Arup Acharya
Renu Tewari
Dinesh Verma

A. Shaikh, R. Tewari, and M. Agrawal, "On the Effectiveness of DNS-based Server Selection," Proc. IEEE INFOCOM 2001, April 2001.PDF. Postscript.

A. Acharya and A. Shaikh, "Using Mobility Support for Request Routing in IPv6 CDNs (R&D Synopsis)", Proc. International Workshop on Web Content Caching and Distribution (WCW '02), August 2002.
Extended version published as IBM Research Report RC 22450 (Postscript) (PDF)

L. Amini, A. Shaikh, and H. Schulzrinne, "Modeling Redirection in Geographically Diverse Server Sets", Proc. of 12th International World Wide Web Conference (WWW 2003), May 2003. (Postscript) (PDF)