Skip to content
Crypto.world

Research

EVM Workloads in the Wild: Evidence for Multi-Dimensional Gas Metering, State Growth, Delayed Execution, and Parallelism

EVM gas metering assumes two things: that the mix of resources a blockchain consumes is stable enough to price everything with one number, and that a transaction behaves at execution time the way it…

EVM Workloads in the Wild: Evidence for Multi-Dimensional Gas Metering, State Growth, Delayed Execution, and Parallelism
arXiv
2606.19869
Published
2026-06-18
Authors
Lioba Heimbach, Kushal Babel, Jason Milionis

AI summary

Overview

Research area: Blockchain systems — EVM execution, gas metering, state management, and Layer-2 scaling (distributed systems measurement).

Technical level: Intermediate. Readers should know roughly what a transaction, gas, and a smart contract are; the paper supplies its own background on EVM internals, so no deep protocol expertise is assumed.

Scope (one sentence): A trace-level measurement study of 2025 Ethereum L1 and Base L2 workloads that decomposes gas into opcode-level resource categories, quantifies persistent state growth, and measures how sensitive transaction execution is to the blockchain state it is evaluated against.

What This Paper Is About

EVM gas metering assumes two things: that the mix of resources a blockchain consumes is stable enough to price everything with one number, and that a transaction behaves at execution time the way it did when it was simulated. The authors measure how badly both assumptions fail on live networks. They do this by breaking down hundreds of millions of transactions on Ethereum and Base into opcode categories and state deltas, and by re-running transactions against older versions of the blockchain state to see how much gas, outcomes, and storage accesses change.

Key Contributions

  1. Trace-level gas decomposition at scale. The authors sample 3,000 blocks per day per chain throughout 2025 — yielding 219,811,453 Ethereum transactions and 256,506,972 Base transactions — and split each transaction into opcode-level execution gas, intrinsic gas, refunds, and persistent state deltas (net storage slots written and net contract bytecode growth). A depth-directed accounting model is used to attribute gas to call-family opcodes, which RPC traces do not expose directly.

  2. Two state-sensitivity datasets. They probe the RPC surface by calling eth_estimateGas at historical states with lookback distances N ∈ {0,1,…,20} and re-execute transactions at lookback distances N ∈ {0,5,10,20} plus the original ("landed") state, recording changes in gas, success/failure, and storage access patterns. This analysis covers all transactions in the sampled blocks during September 2025.

  3. Transaction and address labeling pipeline. Labels from Spellbook (Dune Analytics), DefiLlama, and the Kleros Curate registry cover 46.4% of Ethereum transactions and 19.3% on Base; additional chain-specific steps raise coverage to 62.9% and 75.8% respectively. Labels are mapped to functional categories (MEV, DeFi, Token, Infrastructure, Bridge, NFT, CEX, L2) and used to stratify both workload composition and state sensitivity.

  4. Empirical grounding for multi-dimensional gas metering and state pricing. The measurements are positioned as evidence for pricing distinct resources separately and for explicitly charging for permanent state growth rather than treating it as a transient cost.

Main Findings

  • Execution is a much larger share of gas on Base. Execution gas accounts for 93.30% of total gas on Base versus 79.82% on Ethereum. Mean gas per transaction is 232,936 on Base versus 103,284 on Ethereum (execution component: 217,340 versus 82,441).

  • The resource mix differs sharply across chains. Base spends 29.23% of execution gas on storage reads and 24.30% on compute; Ethereum spends 34.88% on storage writes. Compute and call-related operations also take a larger share on Base (calls: 12.57% versus 10.18%). Contract creation and account structures drive different profiles: logging is 7.19% on Ethereum versus 3.45% on Base.

  • Cross-contract calls on Base are mostly read-only. STATICCALL accounts for 66.8% of call gas on Base, compared with 22.4% on Ethereum.

  • The mix is unstable on a single chain too. On Ethereum, storage writes were 35.6% of execution gas and compute 18.0% in January 2025; by December these had moved to 32.5% and 23.8%, even as total execution gas roughly doubled. This coincides with Ethereum's gas limit doubling from 30M to 60M during 2025, with no opcode repricing — a drift toward Base's more compute-heavy profile.

  • Base touches more distinct storage. 49.7% of SLOAD operations on Base are cold versus 39.6% on Ethereum. Base averages 58.1 SLOAD accesses per transaction (standard deviation 261.1), while Ethereum averages 21.0 (standard deviation 177.3); Ethereum's 25th percentile is 0.0, reflecting many transactions with no storage interaction. Under EIP-2929 a cold SLOAD costs approximately 2,100 gas and a warm one about 100, so Base's workload concentrates more in a comparatively underpriced component.

  • Persistent state growth is large and compositionally different. State growth reaches 435 GB on Base versus 30 GB on Ethereum. Storage slots account for 61.4% of Base's growth and 50.9% of Ethereum's; bytecode contributes 25.4% and 18.2% respectively. Daily growth varies by up to an order of magnitude.

  • Deletion behavior differs. Slot deletion rates are 34.6% on Ethereum versus 22.5% on Base, implying stronger incentives for storage reusability on Ethereum.

  • Simple transfers matter disproportionately on Ethereum. They are 33.2% of Ethereum transactions and only 6.7% of gas, yet contribute 15.4% of state growth through persistent account creations.

  • Execution outcomes are state-sensitive. Gas estimates change across nearby historical states for 46.0% of transactions on Base versus 13.9% on Ethereum, with MEV and DeFi activity especially sensitive. Base's MEV is dominated by "optimistic MEV" — speculative on-chain probing that often produces no state change — and is more state-sensitive in both the slots accessed and the values written. This challenges the assumption, common in parallel-execution work, that a transaction's access set is known before execution, and it limits how effective pre-declared access lists can be.

  • Scope caveat: the supplied text is truncated partway through Section 5.3, so the fuller state-growth analysis the authors present beyond that point is not covered here. Several quantities — for example a detailed breakdown of which transaction categories drive state growth, or the magnitude of gas-estimate error rather than its frequency — are not reported in the available content.

Methodology in Plain English

The authors ran archival nodes of the reth client for both Ethereum and Base and pulled two kinds of data.

First, for the whole of 2025 they picked 3,000 random blocks per day on each chain — about 42% of Ethereum's roughly 7,200 daily blocks and about 7% of Base's roughly 43,200, chosen because full tracing of every block plus repeated re-execution was computationally infeasible. On Base the sampling fraction is lower but the absolute transaction count is comparable. For each transaction they traced every opcode executed, then sorted opcodes into categories such as storage read, storage write, compute, calls, contract operations, and logging. Since RPC traces fold the gas of nested calls into the parent frame, they computed the cost of call opcodes by subtraction: total frame gas minus child-frame gas minus other opcodes in that frame. Intrinsic gas, which RPC does not break out, was reconstructed from explicit rules (base cost of 21,000, calldata, contract creation, EIP-2930 access lists, EIP-7702 authorization lists), with refunds tracked separately and capped per EIP-3529. They also recorded which storage slots changed and which new accounts and bytecode appeared, and extrapolated aggregate figures proportionally from the sample.

Second, for state sensitivity they focused on September 2025, a window of stable protocol parameters between the Pectra fork in May 2025 and later gas limit changes. For each transaction they compared the "landed state" (the state in which it actually executed, taken from canonical traces via debug_traceBlockByNumber) with "lookback states" at distance N, meaning the state at the end of block b−1−N. Gas estimates were requested through eth_estimateGas because that call is cheap enough to run at 21 lookback points; full opcode-level re-execution via debug_traceCall was limited to four lookback points plus the landed state. Write sets came from state diffs using prestateTracer in diff mode, so they capture net-changed slots rather than every write.

Finally, they attached labels to addresses and transactions so results could be split by application type. Public label sources were supplemented by RPC queries detecting well-known pools and tokens (applied to addresses with at least 100 transactions), explorer APIs (Etherscan, Basescan) for addresses with at least 1,000 transactions, and MEV-specific labeling for the top 1,000 addresses by gas usage on each chain — atomic arbitrage and CEX-DEX contracts on Ethereum, cyclic arbitrage contracts on Base. For Base only, unlabeled contracts with at least 10,000 transactions were sampled up to 100 transactions each; those where more than half the sample produced no state change beyond gas accounting were labeled optimistic MEV. A small number of addresses were labeled manually.

Why This Matters

Impact on research. Prior work on gas pricing either benchmarked opcodes under synthetic inputs or analyzed Ethereum alone. This paper provides live, cross-chain, time-series evidence that the resource mix drifts — across L1 and L2 and within a single chain over a year — which is precisely the condition under which one-dimensional metering misallocates capacity. It also challenges a working assumption in the parallel-execution literature: that a transaction's storage access set is deterministic and knowable in advance. Its findings are consistent with Cabral et al.'s result that gas estimation degrades as state evolves (they found estimates largely valid for about 11 blocks), but show the degradation is heterogeneous by transaction type.

Real-world applications:

  • Fee mechanism design. Concrete motivation for multi-dimensional fee markets, such as the proposals the paper cites (e.g., EIP-8011), rather than manual opcode repricing that has historically lagged workload changes.
  • State-growth policy. Evidence for pricing permanent state separately from transient resources, and for state expiry or pruning designs, given 435 GB of growth on Base versus 30 GB on Ethereum in 2025.
  • Wallet and RPC gas estimation. Knowing that gas estimates shift across nearby states for 46.0% of Base transactions tells wallet builders where to add safety margins and which transaction categories need them most.
  • Parallel and delayed execution engineering. Divergent access patterns across states mean speculative execution may cache the wrong data on retry, and misestimated gas limits can incur direct penalties when charges are levied on the declared limit.

Industry relevance. Base is described as the EVM-compatible optimistic rollup with the largest transaction volume and total value locked among L2s, and it inherits Ethereum's opcode prices unchanged despite a markedly different workload. The finding that L2 MEV is dominated by optimistic MEV — speculative probing that usually changes no state — is a Base-specific phenomenon not prevalent on Ethereum L1, and it has direct implications for anyone operating infrastructure on low-fee, high-throughput chains.

Future Directions

  • Designing and calibrating multi-dimensional fee markets. The paper argues the right relative opcode prices should differ across chains; an open question is how such prices would be set and updated as workload mixes drift, given that Ethereum's mix shifted measurably during 2025 with no repricing at all.
  • Explicit pricing of persistent state. With storage slots and bytecode contributing differently to growth on each chain (61.4% versus 50.9% for slots; 25.4% versus 18.2% for bytecode), how should state be charged, and what would change given that slot deletion rates already differ (34.6% on Ethereum versus 22.5% on Base)?
  • Making state-sensitive execution predictable. Whether access lists, gas-limit estimators, or delayed-execution designs can be adapted to state-dependent access patterns — and whether the sensitivity observed in September 2025 holds under higher throughput, encrypted mempools, or the delayed execution models the paper discusses (Monad, Avalanche ACP-194).
  • Extending the measurement. The state-sensitivity analysis is limited to one month, and the state-growth analysis is limited by 3,000-block-per-day sampling (about 7% of Base blocks). Whether the same patterns appear on other L2s and over longer windows is left open.

Target Audience

Protocol researchers and client developers working on gas metering, fee markets, or state management; L2 and rollup engineers inheriting L1 pricing; wallet and RPC infrastructure teams who build gas estimators; and systems researchers studying parallel or delayed execution, who will find the access-pattern variability directly relevant. Graduate students and technically inclined readers looking for a rigorous, data-grounded example of empirical blockchain measurement will also benefit, since the paper supplies its own EVM background.

Authors’ abstract

Gas metering on EVM-compatible blockchains assumes that execution conditions are stable: that the resource mix is constant enough to justify collapsing execution costs into a single scalar with fixed relative prices, and that state drift between submission and execution does not materially alter a transaction's outcome. We measure the extent to which this assumption fails. We present a trace-level measurement study of EVM workloads on Ethereum (L1) and Base (L2) throughout…

The full abstract on arXiv
Read the original paper