Skip multi-dimensional gas accounting for eth_call simulations (~10% latency improvement)

[cristian-dutu]

Is your feature request related to a problem? Please describe.

eth_call simulations spend ~8.4% of total CPU time on multi-dimensional gas accounting (addConstantMultiGas + MultiGas.SaturatingAddInto) that is never consumed. The UsedMultiGas struct tracks gas across 7 Arbitrum resource kinds (Computation, StorageAccess, HistoryGrowth, etc.), but eth_call never produces receipts — the only consumer of this data (state_processor.go → receipt.MultiGasUsed).

This was identified via perf profiling under 100 RPS eth_call load on Arbitrum mainnet:

% CPU	Function	Category
5.9%	addConstantMultiGas	Per-opcode constant multi-gas
2.5%	MultiGas.SaturatingAddInto	Per-opcode dynamic multi-gas
8.4%	Total	Write-only bookkeeping, discarded for eth_call

Describe the solution you'd like

Add a SkipMultiGas bool field to vm.Config and gate the two hot-path multi-gas accumulation calls in the interpreter loop:

// core/vm/interpreter.go — in the Config struct
SkipMultiGas bool // Skip multi-dimensional gas accounting (RPC-only optimization)

// In Run(), before the main loop:
skipMultiGas = evm.Config.SkipMultiGas

// In the main loop — gate the two hot lines:
if !skipMultiGas {
    addConstantMultiGas(&contract.UsedMultiGas, cost, op)
}
// ... and for dynamic gas:
if !skipMultiGas {
    contract.UsedMultiGas.SaturatingAddInto(multigasDynamicCost)
}

Then set SkipMultiGas: true in the eth_call RPC path (internal/ethapi/api.go → doCall):

return applyMessage(ctx, b, args, state, header, timeout, gp, &blockCtx,
    &vm.Config{NoBaseFee: true, SkipMultiGas: true}, precompiles, runCtx)

Safety:

• Single-dimensional gas (contract.Gas) is completely untouched — gas limits, OOG errors, and gasUsed in responses work identically
• UsedMultiGas is write-only during EVM interpretation; no execution path reads it to make branching decisions
• Block production, consensus, Stylus WASM execution, debug_traceCall, and all other RPCs are unaffected (they don't set SkipMultiGas)
• eth_estimateGas could also benefit if it uses the same code path

Describe alternatives you've considered

1. Make addConstantMultiGas a no-op function pointer — introduces indirect call overhead in the hot loop, worse than a branch
2. Compile-time flag — less flexible, prevents using the same binary for both block production and RPC
3. Skip dynamicGas MultiGas construction entirely — would require changing all dynamicGas function signatures to return uint64 when in skip mode; larger refactor for diminishing returns

Additional context

Benchmarked on Arbitrum mainnet, EPYC 4584PX @ 4.8 GHz, 500-request mix (small to xlarge payloads, from 10KB to 500KB-1MB requests), 100 RPS, 3-minute runs:

Metric	Before	After (SkipMultiGas)	Change
P50	87ms	78ms	-10.3%
P90	131ms	122ms	-6.9%
P99	178ms	173ms	-2.8%
Errors	0	0	—

The change is 3 lines of logic + 1 config field. The skipMultiGas local variable ensures the branch predictor handles the constant check with zero overhead after the first iteration.

Tested on Nitro v3.9.7 (75e084e).

[gligneul]

Thanks for the detailed ticket. Could you share how you produced the load to get to these numbers?