This section examines the role of the Anypoint switch in coordinating rate control. This discussion considers the traffic from the switch's viewpoint: because the switch must control each endpoint's rate independently, it views the traffic to or from each endpoint (the n active set members plus the peer) as distinct flows. (This contrasts with the previous section, in which the peer's view of a connection is a pair of flows, one inbound to the ensemble and one outbound from the ensemble.) The switch merges a fan in of n flows--one from each ensemble member--into the connection's outbound flow to the peer, and splits the peer's inbound flow into a fan out of n flows. As flows split and merge, the switch propagates rate control signals to avoid overflowing any receiver or network path. Transport equivalence implies that end nodes do not change their rate control policies to use Anypoint; the end nodes are not aware that a split or merge is occurring. Instead, it is the switch's responsibility to transform and coordinate these signals to induce the correct local behavior from the sources and produce the desired global outcome.
The switch observes rate control signals flowing through it, and can determine if forwarding a frame violates rate limits to the receiver. It can also send rate control signals to any sender. Flow control signals proactively limit the rate of the source; we assume that the transport allows a receiver to rate-limit a source by advertising a flow window. A switch may manipulate these windows to suit its needs [25]. Congestion signals cause a sender to reactively reduce its rate. For example, if the switch drops a packet, a TCP-friendly sender interprets the event as congestion in the usual fashion.
The policy choice is to determine how the switch uses these rate control signals to respond to observed conditions. But the switch cannot predict how the ALRM will route inbound traffic to the ensemble sinks, or what portion of the bandwidth back to the peer will be needed for each source. In either direction, it may optimistically oversubscribe the windows, conservatively rate-limit senders to avoid any overflow, or select any point on the continuum between these extremes. For example, for inbound traffic it may optimistically advertise the sum of the active set flow windows to the peer, or conservatively advertise the minimum window from any sink. For outbound traffic, it may advertise the peer's full window to each ensemble source, partition it evenly among the sources, or overcommit it to an arbitrary degree.
The conservative approaches may limit connection throughput, while the optimistic approaches may cause the switch to overflow a receiver or network path, forcing it to drop packets. A dropped inbound packet induces the peer to throttle its sending rate to the entire ensemble, even if just one sink overflows. The peer's inability to distinguish among ensemble nodes is fundamental to the Anypoint model; we accept it because we assume that the network and memory within the ensemble are well-provisioned in the common case, and aggregate throughput is more important than bandwidth from the peers when the ensemble is overcommitted. For outbound traffic, congestion on the path to the peer results in a lazy throttling of individual sources in the usual fasion.