Anypoint: Extensible Transport Switching on the Edge Ken Yocum, Darrell Anderson, Jeff Chase and Amin Vahdat Department of Computer Science, Duke University {grant,anderson,chase,vahdat}@cs.duke.edu ∗ Anypoint is a new model for one-to-many communication with ensemble sites ”aggregations of end nodes that appear to the external Internet as a uni ed site. Policies for routing Anypoint traf c are de ned by applicationlayer plugins residing in extensible routers at the ensemble edge. Anypoint differs from previous indirection approaches in that its switching functions operate at the transport layer, at the granularity of transport frames. Anypoint transport switching preserves end-to-end transport behavior for rate control, ordering, and reliable delivery. As in illustrating example, we have built an NFS storage router using a host-based Anypoint prototype. Anypoint generalizes œL4-L7  server switches, commonly used for server load balancing. The Anypoint architecture supports ne-grained indirection for contentbased request routing, resource management, and response merging. Anypoint allows independent, concurrent handling of multiple requests arriving on the same persistent transport connection, enabling a new class of virtualization switches for network storage protocols and other Internet services. In contrast, commercial Web switches are limited to service protocols (e.g., HTTP) and issue each request in a separate transport connection, or process requests on each connection serially. Anypoint enables transparent extension and virtualization of network services based on application-layer (L7) protocols. Indirection at the intermediary hides the ensemble ™s internal structure from the connection peer. Clients send inbound traf c to an Anypoint switch using a virtual IP address (VIP) for the ensemble site. The switch merges outbound traf c for the peer into a single stream with the VIP as the source address. Service-speci c ApplicationLayer Routing Modules (ALRMs) in an Anypoint switch examine, transform, and redirect traf c. In this respect, Anypoint is in the spirit of Active Networks and subsequent proposals for enhanced, extensible router architectures. A key principle of our approach is that transport functions ∗ This work is supported in part by the U.S. National Science Foundation (CCR-00-82912, EIA-9972879, and EIA-9870728), Network Appliance, and Cisco Systems, and by an equipment grant from IBM. such as buffering, reassembly, acknowledgment, and retransmission are handled in an end-to-end fashion by the end nodes. Anypoint communication is reliable, partially ordered, and rate-controlled and complements IP-layer approaches (e.g., Anycast and i3), which support widearea ensembles but provide only best-effort packet delivery. Note that ”in contrast to application-level proxies ” an Anypoint intermediary does not terminate transport connections. Anypoint is the rst general indirection approach for connection-oriented transports that operates at the granularity of frames. Anypoint is designed for advanced IP transports with partial ordering and application-level framing, as proposed by Clark and Tennenhouse over a decade ago. This approach is more powerful and elegant than solutions based on TCP connection migration. These transport capabilities help limit data buffering to port queues and frame reassembly, making our approach amenable to implementation on high-speed switches and routers. The Anypoint prototype uses a nonstandard IP-based transport protocol that is framed and partially ordered. We reuse code from TCP, adding framing support based on a subset of the TCP upper-layer framing (TUF) proposal that we call TUF-lite. Each TUF-lite segment is self-describing and contains an integral number of complete frames. A small amount of code at the socket layer supports TUF sockets with UDP-like messaging semantics (sendto, recvfrom). Anypoint is transport equivalent: end nodes use the same transport code for point-topoint and Anypoint connections. We believe that the Anypoint model is fully compatible with SCTP, an accepted IP transport with framing. The switch is implemented as a loadable FreeBSD kernel module. The contributions of this research are to (1) show that transport frame switching at the network edge is a powerful technique to virtualize and extend Internet services, (2) de ne an extensible framework and mechanisms that enable this processing, and (3) explore the implications for Internet service structure, extensible routers, and IP transport protocols. ACM SIGCOMM Computer Communications Review Volume 32, Number 3: July 2002

/lp/association-for-computing-machinery/anypoint-extensible-transport-switching-on-the-edge-9k8RUY4DI1