feat(mip8): storage perf-regression suite + consume timing report

.github/configs/feature.yaml

@@ -6,3 +6,13 @@ monad_runloop:
   evm-type: eels
   # Like `monad`, but `--monad-runloop` and eestnet chain id `30143`
   fill-params: -m blockchain_test --from=MONAD_EIGHT --until=MONAD_TEN --chain-id=30143 --monad-runloop -k "not invalid_header"
+
+monad_benchmark:
+  evm-type: eels
+  # The MIP-8 perf suite, which lives outside the default collection and
+  # so is named explicitly. Filled at the smallest block gas every case
+  # supports: below 10M the 128-read sweeps cannot fit one iteration in a
+  # transaction. The excluded `page_spread` cases size their work from `m`
+  # rather than the gas budget, so they need ~29M and ~177M whatever is
+  # passed here.
+  fill-params: -m blockchain_test tests/benchmark/stateful/mip8_pageified_storage/ --from=MONAD_NINE --until=MONAD_TEN --chain-id=30143 --monad-runloop --gas-benchmark-values=10 -k "not m4096 and not m1024"

MONAD_RUNLOOP_TESTING.md

@@ -21,6 +21,7 @@ Requirements: docker, ~10 GB disk for the builder image and build
 artifacts, ~6 GB RAM for hugepages.
 
 ```sh
+snap install astral-uv --classic
 git clone --branch main \
     git@github.com:monad-exp/monad-eest-rust-harness.git
 cd monad-eest-rust-harness
@@ -68,6 +69,48 @@ uv run consume direct --input ../fixtures_eestnet \
   across ~14 threads), so budget ~5 vCPUs per worker (`-n N` needs
   roughly `5 * N` cores).
 
+## MIP-8 perf-regression tests
+
+`tests/benchmark/stateful/mip8_pageified_storage/test_perf_regression.py`
+fills SLOAD/SSTORE workloads at both forks and times block execution on
+the runloop to compare MONAD_NINE (slot-encoded) vs MONAD_TEN
+(page-encoded).
+
+### Setup
+
+```sh
+sudo tee /etc/sysctl.d/99-benchmark.conf >/dev/null <<'EOF'
+kernel.randomize_va_space = 0
+kernel.perf_event_paranoid = 1
+vm.nr_hugepages = 3072
+EOF
+sudo sysctl --system
+sudo cpupower idle-set -D 1
+```
+
+### Run
+
+From the repo root:
+
+```sh
+tmux new -s perf 'TAG=v4 scripts/perf_cycle.sh'
+```
+
+Fills once, consumes `RUNS` times, and writes the NINE-vs-TEN table to
+`../timing_${TAG}_<utc>_table.md` (headed with the
+cycle time and the four repo SHAs). Knobs:
+
+- `TAG` (required) names every artifact; use a fresh one per experiment.
+- `RUNS` consume passes (default 13), `REPEATS` page-disjoint copies per
+  fixture (cold samples reduced to a `min` within each pass).
+- `BLOCK_GAS_M=N` sets the block gas budget in millions, passed through
+  as `--gas-benchmark-values`; the default 200 matches the gas the
+  runloop stamps. Fixtures land under `for_{fork}_at_0200M/`, so the
+  budget a fixture was built with is visible in its path.
+  `SKIP_FILL=1` reuses an existing `../fixtures_${TAG}`.
+
+Run on a quiet host; timings are noisy under contention.
+
 ## Behavior and known limits
 
 - Fixtures containing expected-invalid blocks are skipped: the ledger

docs/running_tests/consume/direct.md

@@ -8,6 +8,8 @@ uv run consume direct --bin=<evm-binary> [OPTIONS]
 
 - `--bin EVM_BIN`: Path to an evm executable that can process `StateTestFixture` and/or `BlockTestFixture` formats.
 - `--traces`: Collect execution traces from the evm executable.
+- `--timing-report`: Write per-block execution timing to `timing_consume.csv` (see [Block Execution Timing](#block-execution-timing)).
+- `--timing-report-dir DIR`: Directory for that CSV; defaults to the HTML report's directory.
 
 !!! warning "Limited Client Support"
 
@@ -35,6 +37,24 @@ uv run consume direct --bin=<evm-binary> [OPTIONS]
 - **Module scope**: Tests EVM, respectively block import, in isolation, not full client behavior.
 - **Interface dependency**: Requires client-specific test interfaces.
 
+## Block Execution Timing
+
+`--timing-report` writes a `timing_consume.csv` holding the per-block
+measurements the client reported while executing each fixture, one row per
+block.
+
+The `test`, `params` and `fork` columns identify the fixture; the remaining
+columns are whatever metrics the client reported, so they vary by client.
+
+!!! note "Reporting clients only"
+
+    A client only appears in the report if its fixture consumer measures
+    block execution.
+
+!!! warning "Run the timed pass serially"
+
+    `--timing-report` is rejected together with xdist (`-n`).
+
 ## Example Usage
 
 Only run state tests (by using a mark filter, `-m`) from a local `fixtures` folder with go-ethereum:

docs/running_tests/useful_pytest_options.md

@@ -148,3 +148,9 @@ Print relevant test stage timings such as client start-up, payload response time
 ```bash
 uv run consume engine --input=<fixture_input> --timing-data
 ```
+
+Write the per-block execution timing reported by the client to a CSV (`consume direct` only, see [Block Execution Timing](./consume/direct.md#block-execution-timing)):
+
+```bash
+uv run consume direct --input=<fixture_input> --bin=<evm-binary> --timing-report
+```

packages/testing/pyproject.toml

@@ -102,6 +102,7 @@ groupstats = "execution_testing.cli.show_pre_alloc_group_stats:main"
 extract_config = "execution_testing.cli.extract_config:extract_config"
 compare_fixtures = "execution_testing.cli.compare_fixtures:main"
 benchmark_parser = "execution_testing.cli.benchmark_parser:main"
+perf_regression = "execution_testing.cli.perf_regression:main"
 
 [tool.setuptools.packages.find]
 where = ["src"]