Transport

QUIC (Quinn + rustls), TLS 1.3 mandatory with mutual auth pinned to Ed25519 node identities. Nothing is readable on the wire. Every figure below is the node's real [transport] config or a real loopback measurement.

How data actually moves between machines — QUIC, a fast encrypted internet transport (TLS 1.3, mutual auth) — and the knobs that trade latency for throughput.

Impact: Results stream quickly and privately; right-sizing the flow-control window and using parallel streams is what lets a long-distance link reach full speed.

Transport

QUIC

Quinn + rustls

Security

TLS 1.3

mTLS · Ed25519

Congestion

cubic

control algorithm

0-RTT

disabled

session resumption

Best throughput

9 MB/s

measured loopback

Best parallelism

×4

124,869 rows/s

Measured loopback benchmark

Real result-stream throughput vs. parallelism — 80,000 rows, single host, no network. Throughput jumps from parallelism 1→2 then plateaus: real bandwidth-delay-product behavior on loopback.

best p50

4ms

peak

9 MB/s

Parallelism	Rows / sec	MB / sec	p50
×1	59,567	4.3	4ms
×2	121,871	8.7	4ms
×4best	124,869	9	4ms
×8	124,102	8.9	4ms

Bench command

reproduce locally

cargo test -p p2p-node --test benches -- --nocapture

P2P_BENCH_ROWSP2P_BENCH_PARALLELISMP2P_BENCH_COMPRESSIONP2P_BENCH_CONGESTION

Real measurements from cargo test --test benches on this machine (engine mock-1).

Tuning

Per-link transport knobs under [transport]. Every control is seeded from this node's real configured defaults; changes are previewed live below.

Congestion control

config default: cubic

Wire compression

config default: none · min 64.0 KB

GSO offload

GRO offload

Pacing

0-RTT resumption

max uni-streams256

send window64.0 MB

BDP-derived flow control

config default: off — size windows from bandwidth × RTT

Result-stream parallelism1 stream

concurrent unidirectional QUIC streams per call · config default ×1

BDP bandwidth1,000 Mbps

BDP RTT50ms

Estimate

derived from the config on the left

Estimated max throughput

0.40Gbps

400 Mbps · 1 streams · no compression

Flow-control window (fixed send window)

64.0 MB

send_window_bytes (BDP off)

cc=cubic comp=none streams=1
gso=on gro=on pacing=on
0rtt=off bdp=off

Sim seeded from the node's real configured defaults. Measured throughput is the loopback benchmark above (9 MB/s); this estimate models a WAN link.

What's tunable under [transport]

Every knob is config-driven and most are overridable per call.

Knob	Config field	Description
Wire compression	compression.algorithm	none / lz4 / zstd — trade CPU for bytes on the wire (over min_size)
Congestion control	quic.congestion	bbr · cubic · newreno, with optional packet pacing
GSO / GRO offload	quic.gso · quic.gro	batch UDP datagrams through the NIC to cut per-packet syscall cost
Result-stream parallelism	result.parallelism	fan results across multiple uni-streams once over parallel_min_bytes
BDP flow-control target	quic.bdp	derive send/receive windows from bandwidth × RTT instead of fixed bytes
0-RTT / session resumption	quic.enable_0rtt	skip a full handshake RTT on reconnect to a known peer

Per-worker latency (plotly box)

Distribution of commit latencies per worker across the run.

loading plot…

Throughput vs parallelism (plotly)

Measured loopback transfer — doubles 1→2, then plateaus (BDP).

loading plot…