# On the Bandwidth Consumption of Blockchains

Crypto.world

> Blockchain proposals have proliferated, and the network traffic they generate directly raises the cost of hosting nodes — sometimes making high-performance validators economically unviable.

**arXiv**: 2601.18400

**Published**: 2026-01-26

**Authors**: Andrei Lebedev, Vincent Gramoli

## AI summary

## Overview

**Research area:** Distributed systems and blockchain networking — specifically empirical benchmarking of peer-to-peer network traffic in layer-1 blockchain protocols.

**Technical level:** Intermediate. Readers need basic familiarity with blockchain consensus and networking concepts (validators, gossip, WebSockets, TCP/UDP), but no deep protocol expertise.

**Scope:** The paper presents the first empirical comparison of bandwidth consumption across five layer-1 blockchains — Algorand, Aptos, Avalanche, Redbelly and Solana — measured pairwise between nodes under varying network sizes, validator counts, and transaction rates.

## What This Paper Is About

Blockchain proposals have proliferated, and the network traffic they generate directly raises the cost of hosting nodes — sometimes making high-performance validators economically unviable. Yet, according to the authors, no comparative study of blockchain bandwidth consumption existed before this work. The paper fills that gap by deploying five protocols in a controlled 25-VM testbed and measuring, packet by packet, how much traffic each node sends to and receives from every other node.

## Key Contributions

1. **The first empirical comparison of blockchain bandwidth consumption**, measuring five layer-1 protocols (Algorand, Aptos, Avalanche, Redbelly, Solana) rather than comparing them only on throughput or latency.
2. **A fine-grained pairwise measurement methodology**, using `iptables` accounting chains and Stabl observer processes to capture per-peer sending (TX) and receiving (RX) traffic without instrumenting the blockchain software itself.
3. **A structural analysis of traffic skewness**, profiling how bandwidth is distributed across routes (node-to-node pairs) and over time (Setup, Workload, and Cooldown phases).
4. **A scalability classification of blockchains into three bandwidth categories** — node-dependent (Solana), validator-dependent (Algorand, Redbelly), and hybrid (Avalanche, Aptos) — based on how consumption responds to increasing node and validator counts.

## Main Findings

- **Solana consumes the most bandwidth.** With 20 nodes and 20 validators, Solana's heatmap peaks at 17,812 KiB, versus 11,240 KiB for the second-highest, Algorand — a 58% gap attributed to Solana's use of erasure coding and redundant dissemination (Turbine over UDP, Quic for transactions).
- **Transaction size is not the driver.** Aptos transactions are 310 B versus Solana's 215 B (Aptos is 44% larger), and Algorand's are 250 B — larger than Solana's — yet both consume less total bandwidth. Redbelly's transactions are 105 B and Avalanche's 117 B. The dominant factors are the transport protocol (polling vs. WebSockets) and the block propagation strategy (full block download vs. hash comparison).
- **Aptos has high idle overhead from polling.** During the 100 s Setup phase, Aptos nodes transmitted 218,856 KiB while Avalanche transmitted 298 KiB and Redbelly 11 KiB. Aptos clients poll repeatedly even when blocks are empty, reaching roughly 227,000 KiB in the Cooldown phase.
- **Full-block download amplifies egress.** Redbelly received 7,314 KiB of transaction data but transmitted 66,511 KiB back to clients, because clients download the entire block body merely to confirm finality. Solana, which lets clients compare hashes or signatures against locally stored payloads, shows a far more balanced profile (RX 13,099 KiB vs. TX 16,325 KiB).
- **Solana balances load well; Aptos does not.** Aptos concentrates traffic on the five transaction-receiving nodes, which also exchange heavily among themselves (the yellow 5-by-5 sub-matrix) because Quorum Store requires them to collect signatures into a "proof-of-store." Solana shows no standout row or column, indicating a highly redundant but evenly distributed "flood" propagation.
- **Role segregation reduces traffic.** Algorand, Aptos, Avalanche and Redbelly show a dense validator block with little or no traffic between non-validator nodes; in Avalanche's and Redbelly's 10-node/5-validator matrices the bottom-right 5×5 sub-matrix is largely empty. Solana is the outlier: it runs a full mesh where every node, regardless of role, exchanges data with every other node.
- **Bandwidth scales differently by protocol.** Solana's total consumption rises with node count (879,268 KiB at 5 nodes/5 validators to 4,604,375 KiB at 20 nodes/5 validators) but barely moves with validator count (4,604,375 KiB to 4,891,524 KiB at 20 validators). Algorand and Redbelly scale with validators instead — Algorand goes from 287,032 KiB (20 nodes, 5 validators) to 3,681,191 KiB (20 nodes, 20 validators), over 12 times more; Redbelly jumps from 72,098 KiB to 1,084,185 KiB. Aptos and Avalanche grow with both.
- **Higher TPS means lower total bandwidth for a fixed workload.** With the workload fixed at roughly 20,000 transactions and duration scaling inversely from 200 s to 40 s, every protocol consumes less total bandwidth at higher TPS, indicating substantial time-dependent background traffic (heartbeats, empty block proposals, consensus maintenance). Solana drops from about 4.7 GiB (4,905,820 KiB) at 100 TPS to 2.3 GiB (2,387,814 KiB) at 500 TPS; Aptos falls to as little as 283 MiB (289,834 KiB).
- **Magnitude differences are extreme at small scale.** In the 5-node all-validator configuration, Redbelly averages about 1,700 KiB and Avalanche about 2,800 KiB per link, Algorand about 6,500 KiB, while Solana cells consistently exceed 42,000 KiB — nearly 25 times Redbelly for the same workload.

## Methodology in Plain English

The authors built a black-box measurement harness rather than modifying the blockchain clients. They ran a 25-VM cluster (Ubuntu 24.04.1 LTS on a Proxmox cluster of physical servers, each with 4x AMD Opteron 6378 16-core CPUs at 2.40 GHz, 256 GB RAM and 10 GbE NICs), splitting it into 5 client VMs and 20 node VMs, and deployed Algorand v3.27.0, Aptos v1.25.1, Avalanche C-Chain v1.12.1, Redbelly v0.36.2 and Solana Agave v2.0.20.

For each observed node they installed `iptables` rules — one per peer — matching on source or destination IP address, and a monitoring script periodically read the byte counters and reset them to zero. That yields a time series of average TX and RX rates per directed link, which they render as heatmaps of an *M(i,j)* matrix.

Each run followed a fixed timeline: 100 s of Setup with no transactions, 100 s of Workload, and 100 s of Cooldown. Five clients sent transactions to the first five validators, splitting the target load evenly (for example 40 TPS each for 200 TPS total). They varied *N* (total nodes) and *V* (validators) over {5, 10, 15, 20} with *V ≤ N*, and separately swept target TPS from 100 to 500 while holding the workload near 20,000 transactions. The VM specification was deliberately modest and uniform — lower than some of the blockchains recommend — because a single configuration is necessary for a fair comparison and strict hardware requirements on remote nodes are hard to enforce.

## Why This Matters

**Impact on research.** Previous benchmarking frameworks such as Blockbench, Hyperledger Caliper and Gromit prioritize execution capacity (TPS) and treat the network layer as secondary. This work isolates bandwidth as a first-class metric and empirically confirms a gap between theory and practice: theoretical literature favors *O(n)* communication like HotStuff over quadratic BFT, but the measurements show redundancy-heavy Solana consuming the most bandwidth while lean designs like Redbelly stay efficient — echoing earlier findings that quadratic complexity can perform better when well balanced across a quadratic number of routes.

**Real-world applications:**

- **Validator infrastructure planning** — operators can estimate egress from network size and validator count before committing to a provider or hardware tier.
- **Cloud cost modelling** — the paper notes that hyperscaler egress tariffs can push a single validator past US$8,000 per month, motivating a shift to bare-metal providers with unmetered allowances.
- **Protocol design** — quantifying the bandwidth/robustness trade-off informs choices like erasure coding, role segregation, and polling versus WebSocket streaming.
- **Client and RPC implementation** — the Redbelly-vs-Solana comparison shows how much egress can be saved by verifying hashes and signatures instead of downloading full blocks.

**Industry relevance.** Bandwidth is a variable operational cost that scales with participation, not a one-time capital expense. Any operator, exchange, or staking service running nodes at scale can use these classifications to predict how costs change as a network grows.

## Future Directions

- **Extend beyond five protocols and 20 nodes.** The 25-VM setup is justified by prior work showing small-scale controlled experiments reproduce geo-distributed trends, but validating the three-category classification on larger, geographically distributed deployments remains open.
- **Reconcile theoretical communication complexity with measured traffic.** The paper observes that *O(n)* bounds do not predict real-world bandwidth, leaving room for models that account for redundancy, erasure coding, and transport choices.
- **Investigate the Aptos idle-overhead and sharding behaviour.** Aptos shifts from broadcasting to non-validators toward a sparser, possibly sharded or randomized dissemination as validators increase (visible in the 15-node/10-validator configuration); the mechanism and its bandwidth consequences are not fully characterized.
- **Explore bandwidth reduction without losing robustness.** Since Solana's high consumption buys packet-loss tolerance, a natural question is whether that reliability can be preserved at lower cost — and whether polling clients like Aptos can move to streaming.

## Target Audience

Blockchain protocol designers and researchers in distributed systems who want to understand the network-layer costs hidden behind throughput benchmarks; node operators and staking providers estimating infrastructure and egress budgets; and graduate students or engineers entering blockchain performance evaluation, who will find the black-box `iptables` methodology reusable.

## Authors’ abstract

With the advent of blockchain technology, the number of proposals has boomed. The network traffic imposed by these blockchain proposals increases the cost of hosting nodes. Unfortunately, as of today, we are not aware of any comparative study of the bandwidth consumption of blockchains. In this paper, we propose the first empirical comparison of blockchain bandwidth consumption. To this end, we measure the network traffic of blockchain network nodes of five blockchain…

[The full abstract on arXiv](https://arxiv.org/abs/2601.18400)

[Read the original paper](https://arxiv.org/abs/2601.18400)

---

Canonical: https://crypto.world/paper/on-the-bandwidth-consumption-of-blockchains-2601.18400
