Compute Unit Discrepancy vs Mollusk/On-chain Execution

[a-maggi]

Description

When simulating identical Solana transactions, LiteSVM reports ~20% higher compute unit (CU) consumption compared to Mollusk-svm, despite both executing the same instruction sequence successfully. That percentage varies from every transaction. The CU reported by Mollusk is much more accurate to mainet.

Reproduction

Test case: 2-hop Meteora DLMM swap via custom program

Simulator	CU Reported	Delta
Mollusk-svm 0.9.0	90,946	baseline
LiteSVM 0.9.0	109,462	+18,516 (+20.4%)

CU Breakdown Analysis

Component	Mollusk	LiteSVM	Delta
Outer program invoke	87,022	~105,000	+18k
First DLMM swap CPI	41,002	41,002	0
Second DLMM swap CPI	40,879	40,879	0
SPL Token transfers	~12k	~12k	0

The inner CPI calls report identical CU consumption, but the outer program and transaction-level overhead differ significantly.

Root Cause Hypothesis

LiteSVM processes full transactions (send_transaction) while Mollusk processes single instructions (process_instruction). The overhead may come from:

1. Transaction signature verification
2. Account locking/unlocking overhead
3. Transaction metadata tracking
4. Different syscall implementations
5. Fee calculation overhead

[Aursen]

Is it possible to draft me a small snippet to reproduce your metrics?

[naturespowerr]

I also noticed it: #287

When I call Pumpswap CPI from my custom program, it costs me +40000 CU than it is in mainnet runtime.

[procdump]

@a-maggi , @naturespowerr , I'd like to add another hypothesis - on mainnet VMs are jit compiled and hence are probably better optimized. May I ask if you run your tests on Linux or MacOS? Now bear in mind that on MacOS (arm) the SVM runtime defaults to interpreter mode i.e jit isn't supported. On Linux (x86) however it defaults to jit mode and hence it's closer to the mainnet environment. So to confirm this hypothesis can you run your tests on Linux (x86) and share observations?