IP-layer indirection. Anycast [33] and i3 [41] support best-effort packet delivery with indirection at the IP layer. Anypoint operates at the transport layer and above for a related but different purpose. While Anycast is useful for binding to service sites across the wide area, Anypoint enables stateful, reliable, and congestion-aware connections to an ensemble at a logical site.
Web switches and persistent connections. Anypoint is similar in goals and concept to L4-L7 Web server switches, which implement server load balancing (SLB) and/or content-based routing (CBR) for Web server ensembles. The benefits of Web server traffic management are well-established from research studies (e.g., [31]) and commercial experience. Anypoint offers a more powerful mechanism for handling persistent connnections, as in HTTP 1.1, in which a single transport connection carries a stream of requests to simplify congestion control and amortize connection setup costs [28]. Web switches supporting HTTP 1.1 may route all requests on a given connection to a single server, or else they use connection handoff [8,32,37] to migrate the connection to a different server between requests. The first approach does not support CBR and the second forces the server ensemble to process and respond to the requests in strict sequence, as mandated by HTTP 1.1. Anypoint enables independent, pipelined processing of the requests, to generalize redirection to a broader class of service protocols that do not impose this constraint. This is important for fine-grained requests, e.g., storage protocols such as iSCSI or NFS, or RPC protocols such as RMI.
Redirecting proxies. Section 6 compares Anypoint to redirecting proxies that terminate incoming client connections and relay traffic over connections maintained between the proxy and the servers. Proxies do not preserve end-to-end transport semantics because the proxy may acknowledge data before delivering it to the receiver, and this data is lost if the proxy fails. Moreover, the proxy incurs a high runtime overhead to perform full protocol processing for both connection endpoints and high memory overhead to buffer connection data. TCP splicing is one technique to reduce the runtime overheads [19,15,39], and is amenable to switch-based implementations [6]. Anypoint uses related techniques to short-circuit protocol processing and rewrite packets to map among multiple sequence spaces. However, the Anypoint model is fundamentally different: an Anypoint intermediary does not terminate transport connections or otherwise interfere with end-to-end transport functions.
Active Networks and extensible routers. Anypoint is related to Active Networks [44], in which dynamic packets carry behavior (capsules) that execute on routers. Like Active Networks, the Anypoint switch architecture defines a powerful new capability for extending switches and routers through simple, clean abstractions, meeting needs that are currently served in an ad hoc fashion. However, in contrast to capsules, Anypoint extensions--ALRMs--are installed by an authorized configuration tool, are relatively constrained in their actions, and have known, bounded resource and interface requirements. Thus they are more closely related to switchlets in the SwitchWare architecture [2]; Anypoint enables such extensions to perform application-layer functions while preserving transport semantics. Another extensible router architecture is Click [29]; Anypoint could run as a new set of transport-layer and application-layer modules within the Click framework. Anypoint does not require or benefit from the richer interfaces of the NodeOS [34] framework for extensible routers, which supports multiple execution environments similar to those offered in a general-purpose operating system.
Group communication and multicast. Anypoint is similar to group communication in that it supports communication with a dynamic ensemble addressed in a unified way. However, the basic Anypoint abstraction is less powerful in that it supports an indirect frame unicast rather than ordered multicast. The Anypoint framework is powerful enough to allow an ALRM to implement ordered multicast using an approach similar to Amoeba [24], in which a single intermediary for each connection acts as a sequencer.
Scalable Reliable Multicast(SRM) [20] is similar to an Anypoint intermediary in that it manipulates sequence numbers and is based on application-level framing. However, SRM and Anypoint have different goals: Anypoint enables application-layer routing while SRM does not, and Anypoint is not limited to a multicast model.
Traffic shaping and congestion management. One function of an Anypoint intermediary is to control sending rates to preserve the traffic balance between multiple senders and receivers. The technique of rewriting transport-layer flow window advertisements is due to Packeteer [25]. Anypoint connections aggregate communication from the ensemble to the client, and can be viewed as a multipoint-to-point session. The issue of maintaining fairness both among ACP connections and among ensemble members on an ACP connection is analagous to work on session fairness across the Internet [26]. Anypoint is similar to transport-layer bandwidth reservation schemes [42] in its use of per-flow state at the network edge. It is also related to schemes for coordinating congestion control across multiple flows with the same source or destination, when those schemes are applied in intermediaries [11,30].
Multihoming. SCTP includes support for multihoming, which is similar to Anypoint in that a single connection may deliver traffic to a site through multiple IP endpoints. Anypoint differs from multihoming in that it does not require these multiple IP endpoints to coordinate through shared memory, and it uses multiple IP endpoints concurrently for the same connection.