Handling long-tail content in a content delivery network (CDN)

Abstract

A content delivery network has at least a first tier of servers. A content delivery method includes, at a first server in the first tier of servers, obtaining a request from a client for a resource. If the resource is available at the first server or at a peer of the first server, then the resource is served to the client from the first server. Otherwise, it is determined whether the resource is popular, and if the resource is determined to be popular, then the first server obtains the resource and the first server serves the resource to the client. If the resource is determined not to be popular, the client is directed to a second server, not in the first tier of servers, and the second server serves the resource to the client. The second server may be in a second tier of servers or it may be an origin server.

Description

FIELD OF THE INVENTION

This invention relates to content delivery, to content delivery networks (CDNs), and to frameworks and systems using CDNs.

DETAILED DESCRIPTION

Glossary

As used herein, unless stated otherwise, the following terms or abbreviations have the following meanings:

• • 1. IP means Internet Protocol.
  • 2. “IP address” means an address used in the Internet Protocol to identify electronic devices such as servers and the like.
  • 3. HTTP means Hypertext Transfer Protocol.
  • 4. URL means Uniform Resource Locator.
  • 5. DNS means Domain Name System.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention can be better understood with reference to the following drawings. The elements of the drawings are not necessarily to scale, emphasis instead being placed upon clearly illustrating the principles of the present invention. Furthermore, like referenced numerals designate corresponding parts throughout the several views.

FIG. 1 depicts a general hierarchical, multi-tiered content delivery network (CDN);

FIG. 2 shows the logical organization of servers into groups or clusters in a CDN;

FIG. 3 depicts a content delivery framework (CDF) using a content delivery network (CDN);

FIG. 4 depicts the operation of a two-level content delivery network in a content delivery framework;

FIG. 5 is a flowchart showing operation of the popularity services of the CDN of FIG. 4;

FIG. 6 is an exemplary data structure for maintaining the popularity data for a particular popularity server; and

FIG. 7 is a flowchart showing processing steps of a multi-tiered content delivery network.

OVERVIEW

The Internet and the so-called World Wide Web (the “WWW”) have become ubiquitous. Thousands or even tens of thousands of so-called content providers (publishers) now use the Internet (and, particularly, the WWW) to provide all sorts of content to tens or even hundreds of thousands of clients all over the world.

In order to offload the job of serving some or all of their content, many content providers now subscribe to so-called content delivery networks (CDNs). Using a CDN, some (or all) of a content provider's content can be served to clients from the CDN (i.e., from one or more servers in the CDN) instead of from the content provider's server(s). In a caching CDN, content that is served may also be cached on some or all of the CDN servers, either before being served or in response to specific requests for that content.

The term content as used herein means any kind of data, in any form, regardless of its representation and regardless of what it represents. Content may include, without limitation, static and/or dynamic images, text, audio content, including streamed audio, video content, including streamed video, web pages, computer programs, documents, files, and the like. Some content may be embedded in other content, e.g., using markup languages such as HTML and XML. Content includes content which is created or formed or composed specifically in response to a particular request. The term “resource” is sometimes used herein to refer to content.

Certain publishers can have large content libraries in which only a small proportion of the content (the so-called “short head”) is popular enough to benefit from serving through a caching CDN, while the majority of the content (the so-called “long tail”) is accessed only occasionally and not generally worth caching (or even serving from an edge server). This situation would be typical for a content publisher with a very large music or video library. Some music content—the popular content—may be regularly requested, whereas other music—the not popular (also referred to as unpopular) content—may be seldom if ever requested.

Content can become popular (by various measures of popularity) or fade into relative obscurity dynamically, so a content library cannot easily be explicitly partitioned. Instead, the CDN tracks popularity of certain content, and selectively migrates content toward the edge (i.e., toward the tier 1 servers) as that content becomes popular.

A CDN may have one or more tiers of servers, organized hierarchically. FIG. 1 depicts a content delivery network 100 that includes multiple tiers of servers. Specifically, the CDN 100 of FIG. 1 shows j tiers of servers, denoted Tier 1, Tier 2, Tier 3, . . . , Tier j. Each tier of servers may comprise a number of servers organized into server groups (sometimes referred to as server clusters). The Tier 1 servers are also referred to as edge servers, and Tier 1 is sometimes also referred to as the “edge” or the “edge of the CDN.” The Tier 2 servers (when present in a CDN) are also referred to as parent servers.

Furthermore, content stored in any of the intermediate tier servers (e.g., tier 2, tier 3, . . . tier j in FIG. 1) may be logically partitioned across two or more servers in the respective tier group. For example, a first portion (or subset) of a content provider's library may be stored on a first parent server in Group 1, a second portion may be stored on a second server in Group 1, and the last third may be stored on a third server in Group 1. Various partitioning methodologies are described in further detail below. In operation, the partitioning of content libraries across servers in an intermediate tier of a CDN provides a type of preemptive load-balancing such that certain parent servers are only responsible for handling a pre-defined subset of a content provider's library. In other words, redirected requests from the client to a parent/intermediate tier are distributed within the tier in accordance with the particular partitioning methodology being implemented.

For example, in the CDN 100 of FIG. 1, Tier 1 has n groups of servers (denoted “Edge Server Group 1”, “Edge Server Group 2”, . . . , “Edge Server Group n”); tier 2 (the parent servers' tier) has m server groups (the i-th group being denoted “Parent Server Group i”); and tier 3 has k server groups, and so on. Preferably each tier has the same number of server groups.

FIG. 2 shows the logical organization/grouping of servers in a CDN of FIG. 1. In the exemplary CDN of FIG. 2, each tier of servers has the same number (n) of server groups. Those of skill in the art will know and understand, upon reading this description, that each server group may have the same or a different number of servers. Additionally, the number of servers in a server group may vary dynamically. For example, additional servers may be added to a server group to deal with increased load on the group.

The servers in a server group may be homogenous or heterogeneous, and each server in a server group may comprise a cluster of physical servers sharing the same name and/or network address. An example of such a cluster is described in and co-owned patent application No. 61/064,339 (titled “Load-balancing cluster”, filed Feb. 28, 2008), the entire contents of which are incorporated herein by reference for all purposes.

Servers in the same tier and the same group are referred to as peers or peer servers.

A typical CDN has only one or two tiers of servers. A CDN with only one tier will have only edge servers, whereas a CDN with two tiers will have edge servers and parent servers. (At a minimum, a CDN should have at least one tier of servers—the edge servers.)

The grouping of servers in a tier may be based, e.g., on their physical or geographical location. For example, a particular CDN may have six groups—four groups of servers in the United States, group 1 for the West Coast, group 2 for the mid-west, Group 3 for the northeast and Group 4 for the south east; and one group each for Europe and Asia.

A particular CDN server is preferably in only one server group.

In general, some or all of the servers in each tier can exchange data with some or all of the servers in each other tier. Thus, some or all of the parent servers can exchange information with some or all of the edge servers. For the sake of simplicity, in the drawings, each tier of servers is shown as being operationally connectable to each other tier. In some CDNs, however, it may be preferable that the servers in a particular tier can only exchange information with other servers in the same group (i.e., with peer servers) and/or with other servers in the same group in a different tier. For example, in some CDNs, the edge servers in edge server group k, can exchange information with each other and with all servers in parent server group k, and so on.

A content provider's/customer's server (or servers) are also referred to as origin servers. A content provider's origin servers may be owned and/or operated by that content provider or they may be servers provided and/or operated by a third party such as a hosting provider. The hosting provider for a particular content provider may also provide CDN services to that content provider.

A CDN may also include a CDN origin/content cache tier which may be used to cache content from the CDN's subscribers (i.e., from the CDN subscribers' respective origin servers). Those of skill in the art will know and understand, upon reading this description, that a CDN can support one or more subscribers, i.e., that a CDN can function as a shared infrastructure supporting numerous subscribers. The CDN origin tier may also consist of a number of servers, and these servers may also be organized (physically and logically) into a number of regions and/or groups. The server(s) in the CDN origin tier obtain content from the subscribers' origin servers, either on an as needed basis (a pull) or in advance (via a push).

A popularity service 102 (described in greater detail below) is associated with one or more of the server groups in one or more tiers. In a presently preferred exemplary embodiment, some of the parent server groups have popularity services 102 associated therewith. Although shown as a separate component of the group, the popularity service 102 may be integrated into one of the servers in the group. In some cases, the popularity service may have its own server, distinct from any of the CDN servers. The terms “popularity service” and “popularity server” are used interchangeable herein.

In operation, when a client requests content that is to be served using a content delivery framework, the client may be served that content from a server in the CDN or, in some cases, from the subscriber/customer's origin server.

A client may be directed to a CDN and/or to a server in the CDN in any manner using any kind of server selector system 104. As understood by those of skill in the art, the server selector system 104 generally operates to direct a client's requests for content to an appropriate server in order for that content to be served to the requesting client. An appropriate server may be one which is close to the client (by some measure of cost) and/or one which is not too heavily loaded. All sorts of conditions may be applied to the term “appropriate”, and all sorts of information and tests, both static and dynamic, may be used to determine an appropriate server. The server selector system 106 may, e.g., include or operate fully or partially in Domain Name Service (DNS) servers, standalone devices, or a combination thereof. For example, the server selector system 106 may comprise a single level DNS server that selects an appropriate server based at least in part on some combination of the location of the requesting client and the load on some or all of the CDN servers. Those of skill in the art will know and understand, upon reading this description, that a client's location in a network such as the Internet may sometimes only be roughly determined, and the term “location of the client” is generally taken to be a network location corresponding to the client's network service provider.

Although shown as a component in the drawings, the server selector 106 may comprise numerous components. For example, some or all of the server selection may be based on anycast routing, and the server selector 106 may then include routers and associated tables.

In a presently preferred embodiment the server selector 106 is an intelligent traffic manager (ITM)/adaptive traffic controller (ATC) such as described in co-pending U.S. patent application Ser. No. 10/259,497, filed Sep. 30, 2002, and titled “Configurable Adaptive Global Traffic Control And Management,” (published as US 2003-0065762 A1); and in U.S. patent application Ser. No. 11/976,648, filed Oct. 26, 2007, titled “Policy-based content delivery network selection,” (collectively the “ITM applications”), the entire contents of each of which have been incorporated herein by reference for all purposes. In some embodiments the server selector 106 may include a “best” or “optimal” server selector such as disclosed in U.S. Pat. No. 6,185,598 titled, “Optimized Network Resource Location,” the entire contents of which are incorporated herein by reference for all purposes. The '598 patent refers to CDN servers as so-called repeater servers, and describes a so-called “Best Repeater Selector (BRS) mechanism”.

FIG. 3 shows a content delivery framework 300 with a two-level hierarchical CDN consisting of a tier of edge servers (Tier 1) and a tier of parent servers (Tier 2). Some or all of the edge servers can communicate with some or all of the parent servers. The edge servers are divided into n edge server groups, and the parent servers are divided into m parent server groups. In a presently preferred embodiment, the value of m is equal to the value of n, i.e., in this presently preferred embodiment there are the same number of edge server groups as there are parent server groups. A CDN origin/content cache tier stores subscriber content which it obtains from the various subscribers' origin servers. At least one of the parent server groups (in the drawing, Group 1) has a popularity service 102-1 associated therewith. Preferably more than one parent server group has an associated popularity service, and more preferably each parent server group has an associated popularity service.

Although shown in a parent tier, the popularity service may located anywhere in the system, including in the edge tier.

The popularity service may be used by certain, though not necessarily all, content. When only certain content uses the popularity service, content should be designated in order to use the popularity service.

Some or all of the edge servers in a group may use a popularity service to manage the long-tail content of various subscribers. Each edge server that uses a popularity service is referred to as being bound to that popularity service. An edge server that is bound to a popularity service is sometimes referred to herein as a “longtail coserver.”

FIGS. 4 and 5 show the operation of the popularity services of the CDN of FIG. 3.

When a client 106 requests content (e.g., using an HTTP GET request), that request is directed (e.g., by server selector 104-1) to an edge server in order for the content to be served to the client. For certain designated content a popularity check is interposed into the fill side of the caching operation. FIG. 4 illustrates the flow of messages and data in the content delivery framework 300, and FIG. 5 is a flowchart show operation of the popularity services of the CDN of FIG. 4. For the sake of this particular explanation, assume that the client's request has been directed to edge server 108. (Those of skill in the art will know and understand, upon reading this description, that the client's initial request may be directed to any tier in the CDN hierarchy, including, e.g., to the parent tier.) This server is selected using the server selection mechanisms 104-1 associated with the CDN, e.g., using one or more DNS servers and selecting an edge server based on such factors as the location of the requesting client, the load on the network, network traffic conditions, CDN policies, subscriber policies, and the like.

A client 104 requests content from an edge server 108 (at 500 in FIG. 5). The request from the client 104 arrives at edge server 108 (S1 in FIG. 4). The edge server 108 checks to see if object is present (locally or on a peer) and fresh (at 502). If so, the edge server 108 serves the object to the client 104 from the cache (S2, 504), obtaining the object from a peer, if necessary.

In some embodiments, the system may distinguish between on-net and off-net peers and same-switch peers. An on-net peer is a peer on the same backbone network; an off-net peer is a peer located on a different backbone network; and a same-switch peer is a peer directly connected to the same switch as the agent performing the check. In some embodiments, the edge server 108 may only look for the object on some of its peers (e.g., only on same-switch peers) (at 502).

If the object is not available on the edge server 108 or on a peer, the edge server 108 ascertains whether this object is served based on its popularity (i.e., whether this object has been designated so that the object's popularity will be used to determine where it will be served from) (at 506). If so, then the request is sent to the popularity service 102 associated with the edge server 108, in this case, to the popularity server for the this group (S3a).

The determination as to whether this object is designated to be served from a different location, depending on its popularity (at 506), may be made based, at least in part, on the name (hostname) used to request the object.

It is preferably to allow for a mix of edge servers, some performing popularity checking (as described above), while others do not. For those that are not running the popularity service, the name (hostname) used to request an object will resolve to a parent server (that may or may not provide popularity services). If the parent server does not provide popularity services, then the content will be obtained by the edge server from that parent server, and the content will be served to the client. On the other hand, if that parent server does provide popularity services, it can determine whether or not the edge server is a Longtail coserver based, e.g., on the IP (Internet Protocol) address of the edge server. For no-coservers, the parent server can handle the request without any popularity processing.

A request for content may be an initial request for an object or it may be a request for another part of an object, the initial part having already been served to the client. If the request is for the first part of the object (at 508), e.g., the request includes a request for the first byte of the resource (i.e., it is not a range request that starts after the beginning of the file), the popularity service 102 determines (as described below) if the object is currently popular. First, the popularity count for the current period is incremented (at 510). Based on its determination, the popularity service 102 returns one of three possible responses to the edge server 108 (S3b):

• • 1. If the object has not reached a first/minimal level of popularity (at 512): the popularity service sends the edge server an instruction (e.g., HTTP 302) to redirect the client's request to the origin server (or to the CDN origin cache) (if origin redirects are enabled) (at 514).
  • 2. If the object's popularity has exceeded the first/minimal level of popularity but has not yet exceeded a second, mid-tier threshold (at 516): the popularity service sends the edge server an instruction (e.g., HTTP 302) to redirect the client's request to a parent server (if mid-tier redirects are enabled) (at 518). Note that the redirect processing may further involve a determination of which particular parent server is responsible for storing the requested content. Such a determination would be based on how the distinct subsets of a content provider's library are partitioned across parent servers in the same intermediate tier. Further note that it is also possible for subsets of a content provider's library to be partitioned across servers residing in different intermediate tiers of a content delivery network.
  • 3. If the object's popularity has exceeded the mid-tier threshold (i.e., the object is popular): The popularity service sends the edge server an instruction to serve the content itself (at 520). In a presently preferred implementation, the popularity service sends the edge server a redirect (HTTP 302) with a “follow me” flag set, to the origin server or, if there is one, to the parent tier.

If the edge server 108 receives a redirect from the popularity service 102 without the “follow me” flag set (cases 1 and 2 above), it simply forwards the redirect to the client 104 (S4a, 522, 524). If the edge server 108 receives a “follow me” redirect, it obtains and caches the resource (at 526) and serves it to the client (at 528).

If the popularity service 102 is unreachable, unresponsive, or returns a status code indicating an error (other than HTTP 404), the object is served out of the edge's cache server (and an alert condition is raised).

Once content has been cached at an edge server, the edge server will send notifications (e.g., in the form of revalidations) to the popularity service every time it gets another request for that content. E.g., with reference to the flowchart of FIG. 5, if the edge server 108 determines (at 502) that it has the requested content (or can obtain it from a peer), then, in addition to serving the content (at 504), it also instructs the popularity server to increment the object's popularity count for the current period (at 530). This process keeps the popularity servers up to date on the relative popularity of content being served in their region.

In presently preferred embodiments, the server selection mechanism 104 does not rendezvous clients to parent servers/caches. In other words, in these embodiments, client requests are always initially directed by the server selection mechanism to an edge server. In these cases, when a request for a resource arrives at a parent server/cache, that request should preferably be served (and filled if necessary) unconditionally (since any request from a client is assumed to be the result of a redirect served by an edge server).

In an embodiment where the server selector 104 can direct client requests directly to parent servers (or to any tier other than the edge tier), a server obtaining a client request may choose to redirect that request, e.g., based on popularity. However, those of skill in the art will know and understand, upon reading this description, that it is advisable to track the redirection of a request to avoid circular and/or infinite redirection. One way to avoid such a problem is to limit the number of levels of redirection (i.e., to limit the number of redirects to follow). In a presently preferred implementation, if no final server is selected after following, e.g., thirty two redirects, an error is issued. In some embodiments, if no final sever is selected after a predefined number of redirects, then the last server reached may be used to serve the content. One way to prevent looping is to use different server names (aliases) or IP addresses when redirecting requests so that a server receiving a request can tell whether or not it is a redirect. Those of skill in the art will know and understand, upon reading this description, that information can be transferred between servers using, e.g., HTTP headers or the like.

Those of skill in the art will know and understand, upon reading this description, that in a multi-tier CDN, the popularity service may be located at any tier, or there may be popularity services at more than one tier.

The middle (Parent) tier is optional.

Step (4a) may reply with content (if popular), or with a redirect to a parent or origin server (if not), in which the client will make another request (5a or 5b) to that tier to obtain the content.

If the request is an HTTP GET request or the like, it is forwarded to the popularity service. HTTP POST requests should always be forwarded directly to the origin, since that is where they will need to be processed, and the response to a POST request should not be cached. It may sometimes be preferable to direct GET requests to a different origin server than POST requests.

While the invention has been described with reference to the HTTP protocol, those of skill in the art will know and understand, upon reading this description, that different and/or other protocols may be used and are contemplated by the inventors. HTTP is described in various documents, e.g., Hypertext Transfer Protocol—HTTP/1.1, RFC 2616, Network Working Group, the entire contents of which are incorporated herein by reference.

Those of skill in the art will know and understand, upon reading this description, that different thresholds may be established for each tier in the CDN. Further, those of skill in the art will know and understand, upon reading this description, that each content item may have its own thresholds associated therewith. In this manner, the system can check all content for popularity, with the default thresholds being zero. In this manner, every request will automatically cause the popularity to exceed the threshold and will cause the content to be cached.

By positioning Popularity Servers regionally (paired with parent cache servers), popularity and cache tiers can be managed independently, on a regional basis. Content that is popular in one region/group may not be popular in another region/group (especially if each region/group corresponds to a geographic and/or political region).

We consider it desirable that rendezvous to popularity servers prioritize so-called “regional” proximity, so that clients within the same region will tend to cast their popularity “votes” within that region and get consistent treatment of popular resources. However, if there are multiple parent cache servers available, there will generally be no attempt to rendezvous particular clients to particular parents.

Defining & Measuring Popularity

In preferred embodiments, popularity of an object/resource is measured based on the number of times that object/resource is requested in various time periods. FIG. 6 is an exemplary data structure for maintaining the popularity data for a particular popularity server. The data structure 600 in FIG. 6 is a so-called tally hash structure.

In preferred embodiments, some or all edge servers are associated with (or bound to) popularity servers. An edge server that is bound to a popularity server is sometimes referred to as a bound Longtail coserver. Each popularity server in the system allocates a tally hash structure 800 per bound Longtail coserver. A configuration provides the number of resource (hash) slots to allocate. For a presently preferred implementation, the number of hash slots is on the order of 100 million slots per coserver. Each slot is divided into a number of time buckets, preferably 16 time buckets, each bucket being represented by, e.g., a 4-bit unsigned integer. Those of skill in the art will know and understand, upon reading this description, that the selection of the size of the value in each time bucket depends on policy decisions about bounds for the popularity thresholds, and for keeping very popular resources at the edge. The size, however, is heavily influenced by a need for compactness. One 8-byte word can store all time buckets for one resource slot, and therefore, 800 MB would be required per property, and five to eight such properties could be managed per popularity server without paging.

Each time bucket represents a time period, preferably a number of seconds.

The mapping of requests/content to slots is based on some function of the object name and perhaps other information associated with the request for the object. Preferably the mapping of objects to slots is based on a hash or message digest function (such as MAD or the like) over the object name (and preferably including some parts of the query string). Each slot may therefore represent one or more resources. Each time a query/request arrives at a popularity server for an object, the hash is computed and the slot in the table 800 (for the appropriate co-server) is determined, and the counts in that slot are used. In event of a hash collision, it is therefore possible that one bucket will be receiving and representing counts for more than one object. Since this result is generally undesirable (since it could result in cache fills and edge caching of unpopular objects), the number of buckets should be chosen to be as large as practical.

Those of skill in the art will know and understand, upon reading this description, that different and/or other data structures may be used to implement the popularity counting. For example, since in most cases the total number of resources is expected to far exceed the number of popular resources, a balanced b-tree may be preferable to a hash table. In addition, it is possible to reduce the size of the hash slot by using only some part of the hash. However, reducing the number of bytes of the hash used can result in more name collisions.

Although described above with respect to popularity, those of skill in the art will know and understand, upon reading this description, that other factors may be used along with (or instead of) popularity to determine whether or not to redirect requests. A rule base may be used to augment and/or override the popularity measures for certain resources. The rules in the rule base may be static or dynamic and may be set by the CDN administrator and/or the subscriber. For example, a subscriber may not want to pay for certain content to be served from the edge, regardless of its popularity, and may set a rule accordingly (this particular result could also be achieved by setting the thresholds for that particular content to prevent it from ever being cached at the edge).

Occasional log mining could be used to look for hash collisions in actual subscriber content libraries, and the hash function and bucket sizes could be tuned as needed.

At each time bucket boundary, the popularity service will logically “rotate” the buckets and zero out the oldest tally data for each object.

Whenever a coserver's enrollment in the popularity service changes (added or dropped, or perhaps hints changed), the data structures are to be updated.

The popularity of a given object may be determined as a weighted sum of its popularity over successive time periods. More recent time periods may be given higher weights.

In order to determine which content is to be managed by the popularity service, the CDN operator and/or the subscriber may specify:

• • The tiers at which the content will be managed—edge, intermediate (i.e., parent), or origin (subscriber's or storage tier). In order to be meaningful, at least one of intermediate and origin service should be enabled.
  • Content that is to be managed based on its popularity, rather than simply always being served from a cache.

There are several reasons why a publisher/subscriber may not want the so-called “long-tail” content served from a caching CDN service, for example:

• • If cache fills are done from the subscriber's origin server and the subscriber pays for cache fill bandwidth, unnecessary fills for unpopular resources increase bandwidth costs without providing any value.
  • Serving an unpopular resource through a CDN cache adds latency (due to the cache fill), and risks forcing actually popular resources out of cache, potentially causing thrashing. The result is lower efficiency and the risk of degraded service.

For related or similar reasons, a CDN provider generally also does not want to serve long-tail content from an edge cache:

• • If cache fills are done from origin storage, unnecessary cache fills consume bandwidth both from the storage and cache systems, increasing (doubling) the internal cost to serve. This lowers efficiency, requiring relatively more bandwidth internally to serve the same content to the outside.
  • The second argument above also applies from the CDN's perspective: a CDN operator wants to minimize latency as well as the risk of thrashing in order to satisfy all of its subscribers.Names, Addresses & Configuration Data

As is well known, each server in a network may be addressed by means of one or more network address (e.g., Internet Protocol or IP addresses). Each server in a network may also be known by one or more names (so-called hostnames—fully qualified domain names). Hostnames may be mapped to one or more IP addresses. A hostname may correspond to (and thus resolve to) more than one server.

A system such as ITM (described in the ITM patent applications mentioned above), allows a kind of hostname (called a supername) to refer to multiple servers, and resolves the supername to a nearby server.

Preferably the server selection mechanism is ITM, and each popularity server will have supername that resolves to reach nearby popularity server.

When a popularity server shares or is co-located with a parent server, the parent server may use the name by which it was addressed to determine whether to direct a request to the popularity service. That is, parent cache servers that provide popularity service may recognize requests that use one of the aliases reserved for popularity requests, and call into the popularity service to make the fill/no fill decision and return a redirect as described above.

As noted earlier, if the server selector mechanism does not send initial requests to non-edge servers, then all parent cache servers must recognize requests that have been redirected and serve the requested resource, filling it from the origin (or another tier), if necessary.

A servers hostnames are also referred to as its aliases. Each Longtail coserver preferably has at least two aliases (three if a parent cache/server tier is used): the published supername, the hostname used for popularity service requests, and (if used) the hostname used for parent cache redirects.

Popularity servers will preferably be reached via an ITM supername, and ITM will monitor for the service's availability across the set of servers. Popularity servers should be reached using real IP addresses, and not virtual IPs, and will not necessarily be redundant within a cluster. Redundancy can be provided by having multiple servers per supername. However, preferably there will be no attempt to synchronize the popularity tallies on popularity servers, with the expected desirable effect of managing popularity separately on a “regional” basis, the granularity being determined by the number and distribution of popularity servers deployed. Should a popularity server fail, this could cause a discontinuity in popularity responses as a new server becomes active for a given edge location, but this may be mitigated (for very popular resources) by periodic background refreshes.

Information about the resources and caching strategy include the following:

• • the expected total number of resources associated with this coserver.
  • the number of buckets used to store hit counts for each resource.
  • the number of seconds that each bucket represents. Every time this interval goes by, the count in the oldest bucket is thrown out and a new bucket is started with a count of zero.
  • when the sum of all buckets for a given resource reaches this number on
  • any popularity server, the parent caches (if any) that use that server will start to cache the resource.
  • when the sum of all buckets for a given resource reaches this number on any popularity server, the edge caches that use that server will start to cache the resource.
  • the hash algorithm to apply to the resource names (optional). If not specified, a default algorithm (e.g. MD5) will be used.
  • the maximum number of this coserver's resources that should be in cache at an edge at any given time. This value may be ignored in certain embodiments.
  • the maximum number of this coserver's resources that should be in cache at a parent at any given time. This value may be ignored in certain embodiments.

Preferably resources should not be pre-expired on parent cache servers, as that will cause unnecessary requests to the origin server or queries to peer caches.

Those of skill in the art will know and understand, upon reading this description, that a decision to serve at a given tier that is based only on popularity counts, will not take into account capacity to serve at that tier—so this scheme could overload an origin server or parent tier if they do not have sufficient capacity. Further, if popularity is measured in terms of absolute thresholds on numbers of requests, and if the library is sufficiently large, this could cause cache thrashing at the parent or edge tiers.

Authentication with the origin cache or server, if needed, should be done by the server that receives the initial request from the client. During processing of a redirected request.

Various documents, including patents and patent applications, have been incorporated by reference into this application. In case of any conflict between an incorporated document and the present application, the present application, including any definitions herein, will control.

Thus is provided a feature that allows a CDN to be responsive to increasing or decreasing “popularity” of content by shaping where in the CDN content is positioned and served from.

FIG. 7 is a flowchart 700 of processing steps associated with a multi-tiered content delivery network.

In step 705, a first server in a first tier of servers obtains a request from a client for a resource. In one embodiment, the resource is available as part of a content provider's library (e.g., the resource may be a digital movie requested from an on-line streaming movie provider).

In step 710, if the resource is not available at the first server or at a peer of the first server, and additionally if it is determined that the resource is popular (using processing and functionality previously described), then the first server obtains the resource and serves it to the client. For example, the first server may obtain the resource from an origin server and/or a parent server (in any intermediate tier between the first server and the origin server).

In step 715, if the resource is determined not to be popular (again, using processing and functionality previously described), the client is directed to a second server in a second tier of servers (e.g., any intermediate tier) distinct from the first tier of servers. In this example embodiment, the second server comprises a first portion (or subset) of the content provider's library. For the purposes of this example, the first portion comprises at least the resource that was requested by the client. Furthermore, the second tier includes at least one other server that comprises a second portion of the content provider's library. According to one example embodiment, the first portion is distinct from the second portion (i.e., the portions/subsets of the library do not overlap, or overlap only minimally).

In another example embodiment, the portions of the content provider's library are logically partitioned across servers in the second tier. For example, the portions/subsets of a content provider's library can be partitioned alphabetically across servers in the second tier based on a naming convention associated with each distinct resource in the content provider's library (e.g., digital movies having titles beginning with A-F are stored on server A in a second/intermediate tier, digital movies having titles beginning with G-R are stored on server B in the second/intermediate tier, and digital movies having titles beginning with S-Z are stored on server C in the second/intermediate tier).

Assume, for example, that a content provider's library comprises multiple content-types (e.g., movies, video games, music, software patches, etc.). Per one example embodiment, it would be advantageous to store a proportionate amount of each content type on each participating intermediate tier server (e.g., 60% movies, 30% video games, 10% software downloads on each participating server in the tier). In such a scenario, if the demand for a particular content-type increases during a given time period (and/or demand for a different content-type declines), then such fluctuations in demand will generally affect each participating server proportionately.

Noted that resources may be partitioned across multiple servers by naming convention (e.g., alphabetically) as well as having proportionally distributed content per server (e.g., by content-type). With this in mind, it should be further noted that logical partitioning of resources/content among intermediate servers is not limited to just alphabetic naming convention segmentation and/or content-type proportionality. It is contemplated that the partitioning may be based on various metrics such as, but not limited to, numeric hash values (e.g., MD5) associated with content/resource URLs, the size or file type of a resource, other identifiers associated with other various naming conventions (e.g., numerical, alphanumerical, proprietary, encrypted/mangled, etc.), the relative size and traffic behavior of various resources in a content provider's library or the content provider's library as a whole, and the like.

In step 720, the step of directing the client to a second server in the second tier is performed in response to a determination of which second server actually stores the first portion. In one example, the server that stores the first portion (or, for that matter, any portion that comprises the requested resource) is identified using a hash function (similar to the processing and functionality previously described).

In step 725, the second server that stores the first portion serves the resource to the client.

The foregoing description has been presented for purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise forms disclosed. Obvious modifications or variations are possible in light of the above teachings. The embodiment or embodiments discussed were chosen and described to provide the best illustration of the principles of the invention and its practical application to thereby enable one of ordinary skill in the art to utilize the invention in various embodiments and with various modifications as are suited to the particular use contemplated. All such modifications and variations are within the scope of the invention as determined by the appended claims when interpreted in accordance with the breadth to which they are fairly and legally entitled.

Claims

1. A method of content delivery in a content delivery network comprising at least a first tier of servers, the method comprising: (A) at a first server in the first tier of servers, obtaining a request from a client for a resource, wherein the resource is available as part of a content provider's library;(B) determining whether the resource is available at the first server or at a peer of the first server;(C) based on and as a result of said determining in (B), if it is determined that the resource is not available at the first server or at a peer of the first server, determining if the resource is popular, wherein the resource is part of an object, and wherein determining whether said resource is popular is based, at least in part, on whether the resource is an initial part of the object;(D) based on and as a result of said determining in (C), if the resource is determined to be popular in (C), then the first server obtaining the resource and the first server serving the resource to the client, otherwise,(E) based on and as a result of said determining in (C), if the resource is determined in (C) not to be popular, directing the client to a second server in a second tier of servers distinct from the first tier of servers,wherein distinct portions of the content provider's library are logically partitioned across servers in the second tier of servers,wherein the second server comprises a first portion of the content provider's library, the first portion comprising at least the resource, and wherein at least one other server in the second tier of servers comprises a second portion of the content provider's library, said second portion of the content provider's library being distinct from said first portion of the content provider's library, andwherein the second tier of servers is any intermediate tier of servers between the first tier of servers and an origin server that stores resources associated with the content provider's library, and(F) the second server serving the resource to the client.

2. The method of claim 1, wherein the resource is available as part of the content provider's library, and wherein distinct portions of the content provider's library are logically partitioned across servers in the second tier of servers.

3. The method of claim 1, wherein determining whether said resource is popular in (C) is based, at least in part, on whether the resource is for a part of the object other than the initial part of the object.

4. The method of claim 3, wherein the initial part of the object has already been served to the client.

5. The method of claim 1, wherein the determining in (C) whether said resource is popular is determined, at least in part, using information obtained from a popularity service running on one or more popularity servers.

6. The method of claim 5, wherein the one or more popularity servers maintain information on relative popularity of content being served.

7. The method of claim 6, wherein the one or more popularity servers maintain information on relative popularity of content being served in a region comprising the first server.

8. The method of claim 6, wherein only certain content uses the popularity service, and wherein content is designated to be served based on popularity in order to use the popularity service.

9. The method of claim 1, wherein whether or not said resource is to be served based on popularity of the resource is based at least in part on whether or not said resource is designated to be served based on popularity of the resource.

10. The method of claim 1 wherein determining whether said resource is popular is based, at least in part, on a dynamic measure of popularity of the resource.

11. The method of claim 1, wherein determining whether said resource is popular in (C) comprises: determining a relative popularity said resource in a region comprising the first server.

12. A machine-readable non-transitory medium storing one or more programs for execution by one or more processing units of a data processing system, the one or more programs including instructions for: (A) obtaining, at a first server in a first tier of servers in a content delivery network (CDN), a request from a client for a resource, wherein the resource is available as part of a content provider's library;(B) determining whether the resource is available at the first server or at a peer of the first server;(C) based on and as a result of said determining in (B), if it is determined that the resource is not available at the first server or at a peer of the first server, then determining if the resource is popular, wherein the resource is part of an object, and wherein determining whether said resource is popular is based, at least in part, on whether the resource is an initial part of the object;(D) based on and as a result of said determining in (C), if the resource is determined to be popular in (C), then the first server (i) obtaining the resource, and (ii) serving the resource to the client, otherwise,(E) based on and as a result of said determining in (C), if the resource is determined in (C) not to be popular, directing the client to a second server in a second tier of servers distinct from the first tier of servers,wherein the second tier of servers is any intermediate tier between the first tier of servers and an origin server that stores resources associated with the content provider's library, andwherein portions of the content provider's library are logically partitioned across servers in the second tier of servers, and wherein the second server comprises a first portion of the content provider's library, the first portion comprising at least the resource, and wherein at least one other server in the second tier of servers comprises a second portion of the content provider's library, said second portion of the content provider's library being distinct from said first portion of the content provider's library; and(F) the second server serving the resource to the client.

13. The machine-readable non-transitory medium of claim, 12, wherein the first server determines whether said resource is popular based on information obtained from a popularity service running on one or more popularity servers.

14. The machine-readable non-transitory medium of claim 13, wherein the one or more popularity servers maintain information on relative popularity of content being served in a region comprising the first server.

15. The machine-readable non-transitory medium of claim 13, wherein only certain content uses the popularity service, and wherein content is designated to be served based on popularity in order to use the popularity service.

16. The machine-readable non-transitory medium of claim 12, wherein whether or not said resource is to be served based on popularity of the resource is based at least in part on whether or not said resource is designated to be served based on popularity of the resource.

17. The machine-readable non-transitory medium of claim 12, wherein determining whether said resource is popular comprises: determining a relative popularity said resource in a region comprising the first server.

18. A computer program product comprising non-transitory computer-readable media and having computer readable instructions stored on said non-transitory computer-readable media, the computer readable instructions including instructions for implementing a computer-implemented method, said method operable in a content delivery network (CDN), said CDN comprising a plurality of computers, each computer comprising hardware including memory and at least one processor, said method comprising: (A) at a first server in the first tier of servers in the CDN, obtaining a request from a client for a resource, wherein the resource is available as part of a content provider's library;(B) determining whether the resource is available at the first server or at a peer of the first server;(C) based on and as a result of said determining in (B), if it is determined that the resource is not available at the first server or at a peer of the first server, determining if the resource is popular, wherein the resource is part of an object, and wherein determining whether said resource is popular is based, at least in part, on whether the resource is an initial part of the object;(D) based on and as a result of said determining in (C), if the resource is determined to be popular in (C), then the first server obtaining the resource and the first server serving the resource to the client, otherwise,(E) based on and as a result of said determining in (C), if the resource is determined in (C) not to be popular, directing the client to a second server in a second tier of servers distinct from the first tier of servers,wherein the second tier of servers is any intermediate tier of servers between the first tier of servers and an origin server that stores resources associated with the content provider's library, andwherein the second server comprises a first portion of the content provider's library, the first portion comprising at least the resource, and wherein at least one other server in the second tier of servers comprises a second portion of the content provider's library, said second portion of the content provider's library being distinct from said first portion of the content provider's library; and(F) the second server serving the resource to the client.

19. The computer program product of claim 18, wherein whether or not said resource is to be served based on popularity of the resource is based at least in part on whether or not said resource is designated to be served based on popularity of the resource.

20. The computer program product of claim 18, wherein determining whether said resource is popular comprises: determining a relative popularity said resource in a region comprising the first server.