Repository navigation
feat(evm): let benchmarks raise the L1 module cache bound - #27
Conversation
The address-keyed LRU is fixed at 4096 modules. A long replay corpus holds more unique contracts than that -- 15,415 over a 1000-block window -- so every pass evicts and recompiles, and a twelve-pass timing protocol pays that cost twelve times. Measured on a 269-block corpus, 44% of the cold-compile pass went to recompiling contracts the cache had already seen. That traffic is a property of the bound, not of execution, and it drowns the signal a cross-engine comparison is trying to read. Reading the bound from DTVM_EVM_MAX_MODULE_CACHE_SIZE lets a benchmark size the cache to its corpus and measure execution instead of eviction. Production is unaffected: the variable is opt-in and the default stays 4096.
|
No benchmark summary available. |
|
No benchmark summary available. |
⚡ Performance Regression Check Results
|
|
Recording why this stayed on the fork, since GitHub auto-closed it as merged when This is a benchmark affordance, not an engine fix, which is why it was split from the two correctness fixes and why those went upstream (DTVMStack#605 for the exceptional-halt gas rule, DTVMStack#604 for the null-deref misclassification) while this one did not. It is not optional for anyone reproducing the 1000-block campaign: the corpus holds 15,415 unique contracts against a default bound of 4096, so without |
1 participant
Benchmark infrastructure, not a correctness fix. Sent separately from the
exceptional-halt gas fix on purpose — different reviewer concern.
The address-keyed L1 module cache is fixed at 4096 modules. A long replay corpus
holds more unique contracts than that — 15,415 over a 1000-block mainnet window
— so every pass evicts and recompiles, and a twelve-pass timing protocol pays
that cost twelve times. Measured on a 269-block corpus, 44% of the cold-compile
pass went to recompiling contracts the cache had already seen.
That traffic is a property of the bound, not of execution, and it drowns the
signal a cross-engine comparison is trying to read. Reading the bound from
DTVM_EVM_MAX_MODULE_CACHE_SIZElets a benchmark size the cache to its corpusand measure execution instead of eviction.
Production is unaffected: the variable is opt-in, the default stays 4096, and an
unparseable value is warned about and ignored.
Branches from
feat/evm-persistent-code-cacheat35d2f13, so the diff is thiscommit alone.