Skip to content
Crypto.world

The Ledger

Nethermind Researchers Test EIP-8141’s Mempool DoS Limits

A Nethermind benchmark found that raising EIP-8141’s unpaid validation cap can delay block processing when failing transactions consume node CPU. Client protections that pause mempool validation during block processing kept a 300,000-gas cap within the tested delay budget, but reduced the share of honest transactions admitted under a flood.

Nethermind Researchers Test EIP-8141’s Mempool DoS Limits

Nethermind researchers ran a DoS campaign against a Nethermind client

The researchers ran a denial-of-service campaign against a Nethermind client with EIP-8141 activated, testing how failing transactions affect block processing. Stefano De Angelis, Daniil Ankushin and Manuel Arto are the authors, all at Nethermind. The test transactions used VERIFY code to consume their validation budget and fail just before approving payment, so nobody paid for the work.

Source ethresear.ch

EIP-8141 frame transactions can use arbitrary EVM code for validation. Before admitting one to its mempool, a node simulates the transaction’s validation prefix, including the first frame that sets who pays. If that check fails, the node has spent CPU without collecting a fee. The cap on this unpaid work, MAX_VERIFY_GAS, is currently 100,000 gas per transaction.

Why the 100,000-gas limit is too low

The authors say many validation schemes do not fit under the current cap. A single-note Groth16 privacy spend needs at least 211,828 gas, even after optimization, while EVM verification of some SPHINCS+ variants costs up to 276,000 gas. The post also lists BLS signatures at about 130,000 to 140,000 gas, an fflonk verifier at about 209,000, and a PLONK verifier at about 298,000. Raising the budget could make privacy transactions, post-quantum signatures and other frame uses practical, but would also let each failing transaction consume more CPU.

To measure that risk, the researchers tested caps from 100,000 to 500,000 gas. Their failing transactions used either repeated Keccak hashing over 4 KB or stuffed secp256k1 signatures, with the last signature failing. They measured block-import delay, block-building delay and the share of honest frame transactions admitted to the mempool. Each benchmark ran at least twice on each machine.

Block-processing limits vary by machine

The tests put one Nethermind node on a single CPU core shared by block processing and validation, representing a case with no spare core to absorb the attack. The three machines included an AMD EPYC 4344P bare-metal system, a GCE c2-standard-8 and a slower GCE n1-standard-8. The researchers used 25 milliseconds as their main delay budget and also reported results against a 50-millisecond budget.

Without in-client protections, a local-builder profile absorbed about 30 million gas per second of validation within the 25-millisecond budget on a median block. On the slowest machine, which the authors use as their closest full-node profile, block import stayed within that budget only up to about 15 million gas per second. The limit fell to about 10 million gas per second on a p90 block.

At 100 failing transactions per second, a 300,000-gas cap represents 30 million gas per second. That load stayed within 25 milliseconds on the two faster machines for median blocks, but saturated the slowest machine. The post reports that Nethermind processes about 100 failing transactions per second from one connection. At that rate, the slowest machine’s 10-million-gas-per-second limit corresponds to the existing 100,000-gas cap.

Protections shift pressure to the mempool

The researchers implemented a protection in the Nethermind client that pauses validation of gossiped frame transactions while the node processes a block. With that protection, import delay fell to about 2 milliseconds or less on every tested machine. A 300,000-gas cap at 100 transactions per second then stayed within the 25-millisecond budget.

The protection does not remove the CPU competition; it changes which work gets priority. Failing transactions compete with honest frame transactions for mempool validation, reducing the proportion admitted during a flood. Under a 100-transaction-per-second flood at a 300,000-gas cap, the slowest machine admitted only 12% of honest frame transactions. The post says the mempool “admits only 12–27% of honest frame transactions under the flood.”

The benchmark therefore separates two constraints: how much unpaid validation a node can process without delaying blocks, and how much honest traffic its mempool can admit while handling an attack. The authors’ results show that a higher cap can fit within their block-delay budget when validation yields to block processing, while leaving fewer resources for mempool admission.

Sources

  1. Original source Ethereum Research ethresear.ch

Related stories

Explore

More articles