Akamai Technologies, Inc.

Akomai Technologies, Inc.

Adrian Soviani

Matthew S Levine

F Thomson Leighton

Ravi Sundaram

Robert D Kleinberg

Daniel M Lewin

Andrew Parker

Rizwan S Dhanidina

Srikanth Thirumalai

Jay Parikh

Hariharan Rahul

Bruce Maggs

Yoav O Yerushalmi

Anne E Lewin, Legal Representative

Arthur Wagner Berger

Rizwan S Dhanldina

Silvina Hanono-Wachman

A method for Internet delivery in a delivery network established at network locations, the delivery network comprising a plurality of content servers for serving resources. The servers include a plurality of subsets, each subset being located at one of a plurality of Internet data centers. For each Internet Protocol (IP) address block from which requests for content resources are expected to be received, the method generates a candidate list of data centers to be used to service the requests. For the IP address block, the method selects at least one of the data centers from the candidate list. The selected Internet data center for the IP address block is written into a network map. In response to a DNS query, the map is used to identify one of the Internet data centers from the candidate list to be used to service a request for a content resource.

Network performance monitoring in a content delivery system

A method for Internet content delivery in a content delivery network established at network locations, the content delivery network comprising a plurality of content servers for serving content resources. The plurality of content servers includes a plurality of subsets of content servers, each subject being located at one of a plurality of Internet data centers. For each Internet Protocol (IP) address block from which requests for content resources are expected to be received, the method generates a candidate list of Internet data centers to be used to service the requests for content resources. For the IP address block, the method selects at least one of the Internet data centers from the candidate list to be used to service the requests for content resources. The selected Internet data center for the IP address block is written into a network map. The selecting step is carried out concurrently for each IP address block from which requests for content resources are expected to be received such that the network map comprises the selected Internet data center for each IP address block. The network map is then provided to a domain name service (DNS) associated with the content delivery network. In response to a DNS query received at the domain name service associated with the content delivery network, the network map is used to identify one of the Internet data centers from the candidate list to be used to service a request for a content resource.

The invention is an intelligent traffic redirection system that does global load balancing. It can be used in any situation where an end-user requires access to a replicated resource. The method directs end-users to the appropriate replica so that the route to the replica is good from a network standpoint and the replica is not overloaded. The technique preferably uses a Domain Name Service (DNS) to provide IP addresses for the appropriate replica. The most common use is to direct traffic to a mirrored web site.

Global load balancing across mirrored data centers

A network architecture or framework that supports hosting and content distribution on a truly global scale allows a Content Provider to replicate and serve its most popular content at an unlimited number of points throughout the world. The inventive framework comprises a set of servers operating in a distributed manner. The actual content to be served is preferably supported on a set of hosting servers (sometimes referred to as ghost servers). This content comprises HTML page objects that, conventionally, are served from a Content Provider site. A base HTML document portion of a Web page is served from the Content Provider's site while one or more embedded objects for the page are served from the hosting servers, preferably, those hosting servers near the client machine. By serving the base HTML document from the Content Provider's site, the Content Provider maintains control over the content.

Network performance monitoring in a content delivery service

An intelligent traffic redirection system performs global load balancing for Web sites located at mirrored data centers. The system relies on a network map that is generated continuously, preferably for the user-base of the entire Internet. Instead of probing each local name server (or other host) that is connectable to the mirrored data centers, the network map identifies connectivity with respect to a much smaller set of proxy points, called 'core' (or 'common') points. A core point is representative of a set of local name servers (or other hosts) that, from a data center's perspective, share the point. To discover a core point, an incremental trace route is executed from each of the set of mirrored data centers to a local name server that may be used by client to resolve a request for a replica stored at the data centers. An intersection of the trace routes at a common routing point is then identified. Thus, for example, the common routing point may be the first common point for the trace routes when viewed from the perspective of the data centers (or the last common point for the trace routes when viewed from the perspective of the local name server). The common routing point is then identified as the core point for the local name server.

Method for generating a network map

An intelligent traffic redirection system performs global load balancing for Web sites located at mirrored data centers. The system relies on a network map that is generated continuously for the user-base of the entire Internet. Instead of probing each local name server (or other host) that is connectable to the mirrored data centers, the network map identifies connectivity with respect to a much smaller set of proxy points, called 'core' (or 'common') points. A core point then becomes representative of a set of local name servers (or other hosts) that, from a data center's perspective, share the point. Once core points are identified, a systematic methodology is used to estimate predicted actual download times to a given core point from each of the mirrored data centers. Preferably, ICMP (or so-called &#x201c;ping&#x201d; packets) are used to measure roundtrip time (RTT) and latency between a data center and a core point. Using such data, an average latency is calculated, preferably using an exponentially time-weighted average of all previous measurements and the new measurement. A similar function is used to calculate average packet loss. Using the results, a score is generated for each path between one of the data centers and the core point, and the score is representative of a file download time. Preferably, the score is generated by modifying an average latency with a penalty factor dependent on the time-weighted average loss function. Whichever data center has the best score (representing the best-performing network connectivity for that time slice) is then associated with the core point.

Method for predicting file download time from mirrored data centers in a global computer network

An intelligent traffic redirection system performs global load balancing. The system uses a network map that is generated in part by extending a 'sparse' IP address data map. In particular, a method of extending an IP address block map begins by defining a set of one or more upper bound block(s). These upper bound blocks are then used to partition a space of IP addresses into subsets or 'territories', wherein each territory represents a largest set of IP addresses to which a piece of mapping data may be extended. The &#x201c;piece&#x201d; of mapping data typically consists of a host (usually a &#x201c;name server&#x201d; identified by the core point discovery process) IP address and some data about that host, namely, a &#x201c;nearest&#x201d; data center or a flag indicating that either &#x201c;no data&#x201d; exists for that host or that the system is &#x201c;indifferent&#x201d; as to which of a set of mirrored data centers the host should be mapped. A unification algorithm partitions the territory of each upper bound block into the largest possible sub-blocks in which a given unanimity criterion is satisfied and extends the mapping data in each such sub-block to all of the territory of that sub-block.

Growing community of inventors

Cambridge, MA, United States of America

Adrian Soviani