Observations

Anypoint offers three advantages relative to the TCP proxy:

• Efficiency. Memory use is bounded independent of traffic rates, and scales with the number of connections independent of the number of active servers. The Anypoint switch also avoids processing overheads to terminate the protocol.
• End-to-end guarantees. Anypoint offers end-to-end reliability. In contrast, a proxy acks data that it has not yet delivered to the receiving end node. Also, note that the proxy's delivery order is the same as Anypoint's, which is not the ordering specified by its transport (TCP).
• Layer integration. Anypoint allows a continuum of redirection policies that consider Layer 4 state, e.g., for speed-sensitive steering during periods of unbalanced load.

TCP splicing is one technique to reduce the runtime overheads for a proxy [19], and is amenable to switch-based implementations. This technique is related to Anypoint's sequence number translations to short-circuit protocol processing. However, the Anypoint transport model is fundamentally different.

  

Figure: Slite latency and switch CPU utilizations as a function of offered load for varying intermediary configurations.

Interestingly, inbound Anypoint flows in our prototype may slow down relative to a TCP proxy as the ensemble size grows, due to an interaction between the transport's congestion control and acknowledgments from the ensemble. The Anypoint switch merges acks from the ensemble nodes and sends cumulative acks to the peer. If the servers return acks out of order, the switch must delay them to avoid inciting a fast-recovery reaction on the peer, causing it to reduce the congestion window (TCP Reno and later presume that duplicate acknowledgments indicate lost data). Delaying these acks can negatively impact the acknowledgment clocking, lowering throughput.

After these experiments we can make a number of observations about desirable features for Anypoint-compatible transports:

• Explicit rate control. Outbound Anypoint flows should share link bottlenecks in a TCP-friendly manner. An outbound flow is an aggregate of n ensemble sources; in our prototype this flow is likely to be more aggressive than a competing TCP flow. To ensure fairness, the switch must coordinate ensemble sources through explicit congestion control signals.

• Selective acks (SACK). The transport must divorce congestion behavior from reliability. Triple-duplicate cumulative acks are common and meaningless as congestion indicators for Anypoint communication.

• Flexible flow control. The switch can optimistically or conservatively manage the flow windows as described in Section 4.3. If the switch conservatively distributes the peer's advertised receive window across the ensemble sources, it should be able to revoke unused window allocations and redistribute them to active sources. Alternatively, ensemble members could bid for the peer's receive window by advertising to the switch the amount of data they wish to send.