Summary

  • Presto injects a payloadless pseudo-segment at the head of a forward-only RMT pipeline when a later stage needs to correct state held earlier. Ordinary match-action logic then performs the backward update as a new forward pass.
  • The pseudo-segment is a transport control artifact, not retransmitted application data. A defensible operating record must keep its provenance and state mutation separate from DMA placement, host notification, socket result and application receipt.

A segment that is meant to be mistaken for a segment

A conventional CPU handler can revisit memory while processing one TCP event. An RMT pipeline cannot. Each packet carries a bounded header vector through fixed stages; mutable state is local to the stage where it lives. If a later stage discovers that an earlier value must change, the hardware cannot simply jump back.

The Presto paper makes that constraint concrete in receive reassembly. When an in-order segment closes the gap to a stored out-of-order interval, the receiver should move next-seq to the end of the now-contiguous region and return window space. But the gap is known only after examining ooo-head, downstream from the stages holding next-seq and avail.

Presto’s answer is not to send the real packet through the whole pipeline again. The later stage mirrors a small synthetic record to the pipeline head. This pseudo-segment carries no application payload. It encodes the contiguous sequence range and is deliberately made to look like an ordinary in-order segment to the existing stages. The familiar logic advances next-seq, reduces avail and clears the out-of-order interval. A circular write dependency has been translated into a forward event.

That is elegant because the fast path remains simple. It is also why observability needs careful nouns. The pseudo-segment is authoritative enough to mutate transport state. It is not evidence that another copy of the user’s bytes crossed the network, entered a socket or reached application code.

Speculation, validation and correction are separate events

Presto uses a related technique before the pseudo-segment appears. For baseline receive-window tracking, it optimistically advances next-seq, then updates avail and performs the definitive out-of-window check downstream. Well-behaved senders make that assumption useful. A sender that exceeds the advertised window turns it into a correction case.

The paper says the offending segment is dropped and the control plane restores the per-connection state. While restoration is pending, a negative avail causes later boundary checks to fail and a zero window pauses the sender. Payload becomes acceptable only after final validation. The speculative update, validation result, exception, restoration and eventual acceptance are therefore distinct receipts, even if one packet caused all of them.

The pseudo-segment has its own retry semantics. If congestion drops it, the out-of-order interval remains recorded and a later trigger injects another one. The authors describe this as eventual consistency. If intervening in-order traffic has already advanced the window, ordinary overlap trimming removes the obsolete prefix. Success is not “the synthetic packet was emitted”; it is that the intended state transition eventually committed without double accounting.

Transport completion is not application completion

Presto further separates the path in hardware. Received payload may be placed into host memory by DMA, including out-of-order bytes. A dedicated Application Notification block tells libPresto the highest contiguous buffer offset ready for consumption. When a real segment closes a gap, that notification may be issued immediately, before the subsequent pseudo-segment commits the reassembly-state merge.

Those events are coordinated, but they are not synonyms. DMA proves bytes were written to an addressed host buffer. Notification proves the transport exposed a contiguous range to libPresto. A socket result records what the API returned. Only an application-specific acknowledgement or committed downstream state can establish what the application actually consumed or did.

Acknowledgement generation carries the same warning. Presto can mirror reassembly state in a pseudo-segment and transform it into a valid TCP ACK. If that synthetic record is dropped, the paper compares the effect with ordinary ACK loss. Even a transmitted ACK remains transport evidence: it does not prove durable storage, successful parsing or a business transaction.

What the prototype establishes

The work, presented at ACM SIGCOMM 2026 under DOI 10.1145/3789240.3829111, shows that TCP retransmission, reassembly, flow control and congestion control can be expressed as match-action operations on the tested design. The public repository fixes a valuable implementation surface for reproduction. The authors’ APNIC account clearly explains the forward-only constraint and synthetic correction.

None of those facts turns a pipeline pass into an application outcome. Performance numbers belong to the paper’s prototype, hardware and workloads. Repository availability is not independent replication, and a conference result is not evidence of production adoption.

The useful operational lesson is narrower and more durable. Record the real segment identity, speculative assumption, downstream validation, correction trigger, pseudo-segment generation and epoch, exact state mutation, retransmission and reassembly state, DMA placement, host notification, socket result, application receipt and rollback. A system may legitimately optimize across those layers. Its audit trail should not collapse them.

Sources