P2P Networking in Ethereum

Ethereum’s peer-to-peer layer is becoming the critical bottleneck as blob counts and data availability requirements scale. GossipSub, the current broadcast protocol, is bandwidth-inefficient and increasingly mismatched to the demands of PeerDAS and distributed block building. Multiple research groups presented alternative approaches at NoConsensus.wtf 2025. (→ NoConsensus.wtf 2025 — Overview)

Current State: GossipSub

GossipSub is a hybrid push/pull gossip protocol:

  • Eager push (mesh links): each node maintains ~8 full-mesh connections to peers, sending full messages immediately.
  • Lazy pull (gossip): other peers receive only announcements; they pull messages on demand.

Bandwidth math at 6 blobs (August 2025):

  • Block size: ~150 KB compressed on P2P.
  • Gossip amplification factor: theoretically 8× (mesh degree), empirically ~16× (fresh data = almost everyone requests it).
  • Outbound per block: ~14 MB. Monthly for a major relay (Ultrasound): ~2 TB actual (vs. ~793 GB theoretical at 8× amplification).

Post-PeerDAS scaling problem (mempirate / Jonas, Chainbound):

  • PeerDAS introduces 2× erasure coding extension on top of existing blob data.
  • At 48 blobs (proposed blob-only fork targets): outbound per block jumps to ~21 TB/month per relay, costing ~$2,000/month on AWS bandwidth alone.
  • All of this must complete within the 4-second attestation deadline.
  • Critical insight: relays and builders bear bulk of this cost, not validators. The PBS pipeline is the actual propagation bottleneck, not the staking set. EPBS would redistribute this cost but doesn’t reduce it.
  • Risk: if relay costs exceed benefits, relays may artificially cap blob counts below the protocol target.

RLNC: Random Linear Network Coding (Muriel Medard / MIT / Optimum)

RLNC is an alternative broadcast primitive from information theory:

  • Each node generates random linear combinations of received coded packets and forwards them.
  • Decoding = Gaussian elimination once sufficient equations (coded packets) are received.
  • An erasure code (not error-correcting): optimally exploits whatever bandwidth exists in a network of any topology.

Key properties:

  • Provably optimal: achieves the min-cut throughput for any network graph. Not just order-optimal — exactly optimal.
  • Composable: unlike fountain codes (Raptor, LT), RLNC can be recoded at intermediate nodes. Fountain codes are only optimal for point-to-point; RLNC achieves the min-cut on arbitrary mesh networks.
  • Field size matters: operates over extension fields (e.g., GF(2^8)), avoiding the degenerate root problem of binary-field codes.
  • Pull slightly better than push but both approach optimal.

Measured gains vs. GossipSub (Optimum testnet, ~100 geo-distributed nodes): significant latency reduction across all message sizes. No “amplification explosion” because coding prevents redundant full-message retransmission.

Anonymization extension: RLNC over large fields enables interference alignment — encoded packets can be designed so receivers can only decode data they’re authorized to receive, providing a form of anonymization without routing overhead. See Anonymous Broadcast.

PANDAS: P2P Data Availability Sampling (Michal et al.)

PANDAS is a protocol for running DAS (Data Availability Sampling) under Danksharding within Ethereum’s 4-second consensus window.

Danksharding context:

  • Ethereum’s endgame DA layer: 256×256 cell blob matrix (~32 MB raw, ~140 MB after 2D erasure coding extension).
  • Every row and column reconstructible from 50% of cells.
  • Each cell committed via KZG polynomial commitments — any downloaded cell is cryptographically verifiable.
  • Validators need not download the full blob; instead they sample 73 random cells. If all 73 arrive, the blob is with high probability fully available.

Why GossipSub/DHT fail: too slow. PANDAS uses direct UDP communication.

PANDAS protocol:

  1. Builder generates the blob and distributes cell-subsets directly to validators.
  2. Each validator is deterministically assigned 2 rows + 2 columns based on their cryptographic ID — they know exactly who to contact for missing cells.
  3. Validators reconstruct their assigned rows/columns via P2P fetches, then perform DAS sampling.

Performance (testbed, up to 20,000 nodes): completes blob dissemination + sampling within 2.5 seconds on average, leaving 1.5s margin before attestation deadline. Bandwidth per node: 2–3 MB per block (every 12 seconds).

Status: research prototype; direct path to Danksharding deployment on Ethereum mainnet.

WorldChain: Production L2 Bandwidth Snapshot (Leyton Taylor / Tools for Humanity)

WorldChain (OP Stack L2, Tools for Humanity) publishes rare raw L2 infrastructure data:

  • Block builder throughput: 300 Mbps (hardware-limited, not execution-limited).
  • Block processing latency: 40–60 ms per execution payload (~18M gas average).
  • Gas target: 21M (limit: 42M). EIP-1559 denominator: 1/2.
  • Memory: 2–3 GB active (reth); 31.5 GB system-reported (MAP cache inflation).
  • Disk I/O: ~400–600 read ops/sec; ~5,000 write ops/sec.
  • Transaction backlog: typically 40–60 txs pending; spikes to 250–1,000.
  • TPS: 20–40 typical; occasional spikes.
  • Inbound requests to sequencer endpoint: 1,000–2,000/sec (includes health checks).
  • ~40–50% of volume via ERC-4337; split roughly 50/50 between World app-native and MEV/peripheral activity.

Connections

Open Questions

  • Will GossipSub be modified (e.g., EF’s WFR gossip proposal) before PeerDAS ships, or will bandwidth costs force relay centralization?
  • At what blob count does relay economics tip toward an artificial cap below the protocol target?
  • Can RLNC be deployed as a GossipSub upgrade without protocol-level changes to Ethereum?
  • Does PANDAS require protocol changes, or can it be implemented as a client-layer P2P optimization?