feat: set up codspeed (#13372)

.github/scripts/codspeed-build.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+# TODO: Benchmarks run WAY too slow due to excessive amount of iterations.
+
+cmd=(cargo codspeed build --profile profiling)
+excludes=(
+    # Unnecessary
+    --exclude reth-libmdbx
+    # Build is too slow
+    --exclude reth-network
+    # Built separately 
+    --exclude reth-transaction-pool
+    # TODO: some benchmarks panic: https://github.com/paradigmxyz/reth/actions/runs/12307046814/job/34349955788
+    --exclude reth-db
+    --exclude reth-trie-parallel
+    --exclude reth-engine-tree
+)
+
+"${cmd[@]}" --features test-utils --workspace "${excludes[@]}"
+
+# TODO: Slow benchmarks due to too many iterations
+## "${cmd[@]}" -p reth-transaction-pool --features test-utils,arbitrary

.github/workflows/bench.yml

@@ -1,4 +1,4 @@
-# Runs benchmarks on serialization/deserialization of storage values and keys.
+# Runs benchmarks.
 
 on:
   pull_request:
@@ -9,7 +9,6 @@ on:
 env:
   CARGO_TERM_COLOR: always
   BASELINE: base
-  IAI_CALLGRIND_RUNNER: iai-callgrind-runner
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
@@ -17,47 +16,26 @@ concurrency:
 
 name: bench
 jobs:
-  iai:
+  codspeed:
     runs-on:
       group: Reth
-    # Only run benchmarks in merge groups and on main
-    if: github.event_name != 'pull_request'
     steps:
       - uses: actions/checkout@v4
-      - name: Install Valgrind
+        with:
-        run: sudo apt update && sudo apt install valgrind
+          submodules: true
+      - uses: rui314/setup-mold@v1
       - uses: dtolnay/rust-toolchain@stable
       - uses: Swatinem/rust-cache@v2
         with:
           cache-on-failure: true
-      - name: Install cargo-binstall
+      - name: Install cargo-codspeed
-        uses: taiki-e/install-action@cargo-binstall
+        uses: taiki-e/install-action@v2
-      - name: Install iai-callgrind-runner
-        run: |
-          echo "::group::Install"
-          version=$(cargo metadata --format-version=1 |\
-            jq '.packages[] | select(.name == "iai-callgrind").version' |\
-            tr -d '"'
-          )
-          cargo binstall iai-callgrind-runner --version $version --no-confirm --no-symlinks --force
-          echo "::endgroup::"
-          echo "::group::Verification"
-          which iai-callgrind-runner
-          echo "::endgroup::"
-      - name: Checkout base
-        uses: actions/checkout@v4
         with:
-          ref: ${{ github.base_ref || 'main' }}
+          tool: cargo-codspeed
-      # On `main` branch, generates test vectors and serializes them to disk using `serde-json`.
+      - name: Build the benchmark target(s)
-      - name: Generate test vectors
+        run: ./.github/scripts/codspeed-build.sh
-        run: cargo run --bin reth --features dev -- test-vectors tables
+      - name: Run the benchmarks
-      # Runs iai and stores `main` baseline report for comparison later on $BASELINE.
+        uses: CodSpeedHQ/action@v3
-      - name: Save baseline
-        run: cargo bench -p reth-db --bench iai --profile profiling --features test-utils -- --save-baseline=$BASELINE
-      - name: Checkout PR
-        uses: actions/checkout@v4
         with:
-          clean: false
+          run: cargo codspeed run --workspace
-      # Runs iai on incoming merge using previously generated test-vectors and compares the report against `main` report.
+          token: ${{ secrets.CODSPEED_TOKEN }}
-      - name: Compare PR benchmarks
-        run: cargo bench -p reth-db --bench iai --profile profiling --features test-utils -- --baseline=$BASELINE

.gitignore

@@ -19,7 +19,7 @@ target/
 testdata/micro/db
 
 # Generated data for stage benchmarks
-crates/stages/testdata
+crates/stages/stages/testdata
 
 # Prometheus data dir
 data/

Cargo.lock

@@ -1843,6 +1843,30 @@ version = "0.7.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "afb84c814227b90d6895e01398aee0d8033c00e7466aca416fb6a8e0eb19d8a7"
 
+[[package]]
+name = "codspeed"
+version = "2.7.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "450a0e9df9df1c154156f4344f99d8f6f6e69d0fc4de96ef6e2e68b2ec3bce97"
+dependencies = [
+ "colored",
+ "libc",
+ "serde_json",
+]
+
+[[package]]
+name = "codspeed-criterion-compat"
+version = "2.7.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8eb1a6cb9c20e177fde58cdef97c1c7c9264eb1424fe45c4fccedc2fb078a569"
+dependencies = [
+ "codspeed",
+ "colored",
+ "criterion",
+ "futures",
+ "tokio",
+]
+
 [[package]]
 name = "coins-bip32"
 version = "0.12.0"
@@ -1900,6 +1924,16 @@ version = "1.0.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5b63caa9aa9397e2d9480a9b13673856c78d8ac123288526c37d7839f2a86990"
 
+[[package]]
+name = "colored"
+version = "2.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cbf2150cce219b664a8a70df7a1f933836724b503f8a413af9365b4dcc4d90b8"
+dependencies = [
+ "lazy_static",
+ "windows-sys 0.48.0",
+]
+
 [[package]]
 name = "combine"
 version = "4.6.7"
@@ -3934,42 +3968,6 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "iai-callgrind"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "22275f8051874cd2f05b2aa1e0098d5cbec34df30ff92f1a1e2686a4cefed870"
-dependencies = [
- "bincode",
- "derive_more",
- "iai-callgrind-macros",
- "iai-callgrind-runner",
-]
-
-[[package]]
-name = "iai-callgrind-macros"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e8e6677dc52bd798b988e62ffd6831bf7eb46e4348cb1c74c1164954ebd0e5a1"
-dependencies = [
- "derive_more",
- "proc-macro-error2",
- "proc-macro2",
- "quote",
- "serde",
- "serde_json",
- "syn 2.0.90",
-]
-
-[[package]]
-name = "iai-callgrind-runner"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a02dd95fe4949513b45a328b5b18f527ee02e96f3428b48090aa7cf9043ab0b8"
-dependencies = [
- "serde",
-]
-
 [[package]]
 name = "iana-time-zone"
 version = "0.1.61"
@@ -6998,10 +6996,9 @@ dependencies = [
  "arbitrary",
  "assert_matches",
  "bytes",
- "criterion",
+ "codspeed-criterion-compat",
  "derive_more",
  "eyre",
- "iai-callgrind",
  "metrics",
  "page_size",
  "parking_lot",
@@ -7391,7 +7388,7 @@ dependencies = [
  "alloy-rlp",
  "alloy-rpc-types-engine",
  "assert_matches",
- "criterion",
+ "codspeed-criterion-compat",
  "crossbeam-channel",
  "derive_more",
  "futures",
@@ -7912,7 +7909,7 @@ version = "1.1.4"
 dependencies = [
  "bitflags 2.6.0",
  "byteorder",
- "criterion",
+ "codspeed-criterion-compat",
  "dashmap 6.1.0",
  "derive_more",
  "indexmap 2.6.0",
@@ -7979,7 +7976,7 @@ dependencies = [
  "alloy-rlp",
  "aquamarine",
  "auto_impl",
- "criterion",
+ "codspeed-criterion-compat",
  "derive_more",
  "discv5",
  "enr",
@@ -8762,7 +8759,7 @@ dependencies = [
  "bincode",
  "bytes",
  "c-kzg",
- "criterion",
+ "codspeed-criterion-compat",
  "derive_more",
  "modular-bitfield",
  "once_cell",
@@ -9287,7 +9284,7 @@ dependencies = [
  "alloy-rlp",
  "assert_matches",
  "bincode",
- "criterion",
+ "codspeed-criterion-compat",
  "futures-util",
  "itertools 0.13.0",
  "num-traits",
@@ -9514,7 +9511,7 @@ dependencies = [
  "assert_matches",
  "auto_impl",
  "bitflags 2.6.0",
- "criterion",
+ "codspeed-criterion-compat",
  "futures-util",
  "metrics",
  "parking_lot",
@@ -9560,7 +9557,7 @@ dependencies = [
  "alloy-rlp",
  "alloy-trie",
  "auto_impl",
- "criterion",
+ "codspeed-criterion-compat",
  "itertools 0.13.0",
  "metrics",
  "proptest",
@@ -9593,7 +9590,7 @@ dependencies = [
  "arbitrary",
  "bincode",
  "bytes",
- "criterion",
+ "codspeed-criterion-compat",
  "derive_more",
  "hash-db",
  "itertools 0.13.0",
@@ -9643,7 +9640,7 @@ version = "1.1.4"
 dependencies = [
  "alloy-primitives",
  "alloy-rlp",
- "criterion",
+ "codspeed-criterion-compat",
  "derive_more",
  "itertools 0.13.0",
  "metrics",
@@ -9671,7 +9668,7 @@ dependencies = [
  "alloy-rlp",
  "arbitrary",
  "assert_matches",
- "criterion",
+ "codspeed-criterion-compat",
  "itertools 0.13.0",
  "pretty_assertions",
  "proptest",

Cargo.toml

@@ -601,8 +601,7 @@ toml = "0.8"
 # misc-testing
 arbitrary = "1.3"
 assert_matches = "1.5.0"
-criterion = "0.5"
+criterion = { package = "codspeed-criterion-compat", version = "2.7" }
-iai-callgrind = "0.14"
 pprof = "0.14"
 proptest = "1.4"
 proptest-derive = "0.5"

crates/blockchain-tree/Cargo.toml

@@ -77,7 +77,8 @@ test-utils = [
 	"reth-db-api/test-utils",
 	"reth-provider/test-utils",
 	"reth-trie-db/test-utils",
-	"reth-trie/test-utils"
+	"reth-trie/test-utils",
+	"reth-trie-parallel/test-utils"
 ]
 optimism = [
 	"reth-primitives/optimism",

crates/engine/tree/Cargo.toml

@@ -120,6 +120,8 @@ test-utils = [
     "reth-static-file",
     "reth-tracing",
     "reth-trie/test-utils",
-    "reth-trie-db/test-utils",
+    "reth-trie-sparse/test-utils",
     "reth-prune-types?/test-utils",
+    "reth-trie-db/test-utils",
+    "reth-trie-parallel/test-utils",
 ]

crates/engine/tree/benches/channel_perf.rs

@@ -7,7 +7,7 @@ use revm_primitives::{
     Account, AccountInfo, AccountStatus, Address, EvmState, EvmStorage, EvmStorageSlot, HashMap,
     B256, U256,
 };
-use std::thread;
+use std::{hint::black_box, thread};
 
 /// Creates a mock state with the specified number of accounts for benchmarking
 fn create_bench_state(num_accounts: usize) -> EvmState {
@@ -47,7 +47,7 @@ impl StdStateRootTask {
 
     fn run(self) {
         while let Ok(state) = self.rx.recv() {
-            criterion::black_box(state);
+            black_box(state);
         }
     }
 }
@@ -64,7 +64,7 @@ impl CrossbeamStateRootTask {
 
     fn run(self) {
         while let Ok(state) = self.rx.recv() {
-            criterion::black_box(state);
+            black_box(state);
         }
     }
 }

crates/engine/tree/benches/state_root_task.rs

@@ -3,7 +3,7 @@
 
 #![allow(missing_docs)]
 
-use criterion::{black_box, criterion_group, criterion_main, BenchmarkId, Criterion};
+use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion};
 use reth_engine_tree::tree::root::{StateRootConfig, StateRootTask};
 use reth_evm::system_calls::OnStateHook;
 use reth_primitives::{Account as RethAccount, StorageEntry};
@@ -22,6 +22,7 @@ use revm_primitives::{
     Account as RevmAccount, AccountInfo, AccountStatus, Address, EvmState, EvmStorageSlot, HashMap,
     B256, KECCAK_EMPTY, U256,
 };
+use std::hint::black_box;
 
 #[derive(Debug, Clone)]
 struct BenchParams {

crates/net/ecies/src/algorithm.rs

@@ -975,6 +975,5 @@ mod tests {
                 &mut dest,
             );
         }
-        std::hint::black_box(());
     }
 }

crates/primitives/benches/validate_blob_tx.rs

@@ -69,12 +69,10 @@ fn validate_blob_tx(
 
     // for now we just use the default SubPoolLimit
     group.bench_function(group_id, |b| {
+        let kzg_settings = kzg_settings.get();
         b.iter_with_setup(setup, |(tx, blob_sidecar)| {
-            if let Err(err) =
+            let r = tx.validate_blob(&blob_sidecar, kzg_settings);
-                std::hint::black_box(tx.validate_blob(&blob_sidecar, kzg_settings.get()))
+            (r, tx, blob_sidecar)
-            {
-                println!("Validation failed: {err:?}");
-            }
         });
     });
 }

crates/stages/stages/benches/criterion.rs

@@ -1,11 +1,10 @@
 #![allow(missing_docs)]
-use criterion::{criterion_main, measurement::WallTime, BenchmarkGroup, Criterion};
+#![allow(unexpected_cfgs)]
-#[cfg(not(target_os = "windows"))]
-use pprof::criterion::{Output, PProfProfiler};
-use reth_config::config::{EtlConfig, TransactionLookupConfig};
-use reth_db::{test_utils::TempDatabase, Database, DatabaseEnv};
 
 use alloy_primitives::BlockNumber;
+use criterion::{criterion_main, measurement::WallTime, BenchmarkGroup, Criterion};
+use reth_config::config::{EtlConfig, TransactionLookupConfig};
+use reth_db::{test_utils::TempDatabase, Database, DatabaseEnv};
 use reth_provider::{test_utils::MockNodeTypesWithDB, DatabaseProvider, DatabaseProviderFactory};
 use reth_stages::{
     stages::{MerkleStage, SenderRecoveryStage, TransactionLookupStage},
@@ -22,25 +21,30 @@ use setup::StageRange;
 // Expanded form of `criterion_group!`
 //
 // This is currently needed to only instantiate the tokio runtime once.
+#[cfg(not(codspeed))]
 fn benches() {
-    #[cfg(not(target_os = "windows"))]
+    #[cfg(not(windows))]
-    let mut criterion = Criterion::default()
+    use pprof::criterion::{Output, PProfProfiler};
-        .with_profiler(PProfProfiler::new(1000, Output::Flamegraph(None)))
-        .configure_from_args();
 
-    let runtime = Runtime::new().unwrap();
+    let criterion = Criterion::default();
-    let _guard = runtime.enter();
+    #[cfg(not(windows))]
-
+    let criterion = criterion.with_profiler(PProfProfiler::new(1000, Output::Flamegraph(None)));
-    #[cfg(target_os = "windows")]
+    run_benches(&mut criterion.configure_from_args());
-    let mut criterion = Criterion::default().configure_from_args();
-
-    transaction_lookup(&mut criterion, &runtime);
-    account_hashing(&mut criterion, &runtime);
-    senders(&mut criterion, &runtime);
-    merkle(&mut criterion, &runtime);
 }
 
+fn run_benches(criterion: &mut Criterion) {
+    let runtime = Runtime::new().unwrap();
+    let _guard = runtime.enter();
+    transaction_lookup(criterion, &runtime);
+    account_hashing(criterion, &runtime);
+    senders(criterion, &runtime);
+    merkle(criterion, &runtime);
+}
+
+#[cfg(not(codspeed))]
 criterion_main!(benches);
+#[cfg(codspeed)]
+criterion_main!(run_benches);
 
 const DEFAULT_NUM_BLOCKS: u64 = 10_000;
 

crates/stages/stages/benches/setup/mod.rs

@@ -88,6 +88,12 @@ where
 // Helper for generating testdata for the benchmarks.
 // Returns the path to the database file.
 pub(crate) fn txs_testdata(num_blocks: u64) -> TestStageDB {
+    // This is way too slow.
+    #[allow(unexpected_cfgs)]
+    if cfg!(codspeed) {
+        std::process::exit(0);
+    }
+
     let txs_range = 100..150;
 
     // number of storage changes per transition

crates/storage/db/Cargo.toml

@@ -69,7 +69,6 @@ pprof = { workspace = true, features = [
     "criterion",
 ] }
 criterion.workspace = true
-iai-callgrind.workspace = true
 
 arbitrary = { workspace = true, features = ["derive"] }
 proptest.workspace = true
@@ -125,11 +124,6 @@ name = "criterion"
 required-features = ["test-utils"]
 harness = false
 
-[[bench]]
-name = "iai"
-required-features = ["test-utils"]
-harness = false
-
 [[bench]]
 name = "get"
 required-features = ["test-utils"]

crates/storage/db/benches/criterion.rs

@@ -3,7 +3,7 @@
 use std::{path::Path, sync::Arc};
 
 use criterion::{
-    black_box, criterion_group, criterion_main, measurement::WallTime, BenchmarkGroup, Criterion,
+    criterion_group, criterion_main, measurement::WallTime, BenchmarkGroup, Criterion,
 };
 use pprof::criterion::{Output, PProfProfiler};
 use reth_db::{tables::*, test_utils::create_test_rw_db_with_path};
@@ -71,12 +71,9 @@ where
         b.iter_with_setup(
             || input.clone(),
             |input| {
-                {
+                for (k, _, _, _) in input {
-                    for (k, _, _, _) in input {
+                    k.encode();
-                        k.encode();
+                }
-                    }
-                };
-                black_box(());
             },
         )
     });
@@ -85,12 +82,9 @@ where
         b.iter_with_setup(
             || input.clone(),
             |input| {
-                {
+                for (_, k, _, _) in input {
-                    for (_, k, _, _) in input {
+                    let _ = <T as Table>::Key::decode(&k);
-                        let _ = <T as Table>::Key::decode(&k);
+                }
-                    }
-                };
-                black_box(());
             },
         )
     });
@@ -99,12 +93,9 @@ where
         b.iter_with_setup(
             || input.clone(),
             |input| {
-                {
+                for (_, _, v, _) in input {
-                    for (_, _, v, _) in input {
+                    v.compress();
-                        v.compress();
+                }
-                    }
-                };
-                black_box(());
             },
         )
     });
@@ -113,12 +104,9 @@ where
         b.iter_with_setup(
             || input.clone(),
             |input| {
-                {
+                for (_, _, _, v) in input {
-                    for (_, _, _, v) in input {
+                    let _ = <T as Table>::Value::decompress(&v);
-                        let _ = <T as Table>::Value::decompress(&v);
+                }
-                    }
-                };
-                black_box(());
             },
         )
     });
@@ -148,14 +136,10 @@ where
                 // Create TX
                 let tx = db.tx_mut().expect("tx");
                 let mut crsr = tx.cursor_write::<T>().expect("cursor");
-
+                for (k, _, v, _) in input {
-                black_box({
+                    crsr.append(k, v).expect("submit");
-                    for (k, _, v, _) in input {
+                }
-                        crsr.append(k, v).expect("submit");
+                tx.inner.commit().unwrap()
-                    }
-
-                    tx.inner.commit().unwrap()
-                });
             },
         )
     });
@@ -171,15 +155,12 @@ where
                 // Create TX
                 let tx = db.tx_mut().expect("tx");
                 let mut crsr = tx.cursor_write::<T>().expect("cursor");
+                for index in RANDOM_INDEXES {
+                    let (k, _, v, _) = input.get(index).unwrap().clone();
+                    crsr.insert(k, v).expect("submit");
+                }
 
-                black_box({
+                tx.inner.commit().unwrap()
-                    for index in RANDOM_INDEXES {
-                        let (k, _, v, _) = input.get(index).unwrap().clone();
-                        crsr.insert(k, v).expect("submit");
-                    }
-
-                    tx.inner.commit().unwrap()
-                });
             },
         )
     });
@@ -190,15 +171,11 @@ where
         b.iter(|| {
             // Create TX
             let tx = db.tx().expect("tx");
-
+            let mut cursor = tx.cursor_read::<T>().expect("cursor");
-            {
+            let walker = cursor.walk(Some(input.first().unwrap().0.clone())).unwrap();
-                let mut cursor = tx.cursor_read::<T>().expect("cursor");
+            for element in walker {
-                let walker = cursor.walk(Some(input.first().unwrap().0.clone())).unwrap();
+                element.unwrap();
-                for element in walker {
+            }
-                    element.unwrap();
-                }
-            };
-            black_box(());
         })
     });
 
@@ -208,14 +185,10 @@ where
         b.iter(|| {
             // Create TX
             let tx = db.tx().expect("tx");
-
+            for index in RANDOM_INDEXES {
-            {
+                let mut cursor = tx.cursor_read::<T>().expect("cursor");
-                for index in RANDOM_INDEXES {
+                cursor.seek_exact(input.get(index).unwrap().0.clone()).unwrap();
-                    let mut cursor = tx.cursor_read::<T>().expect("cursor");
+            }
-                    cursor.seek_exact(input.get(index).unwrap().0.clone()).unwrap();
-                }
-            };
-            black_box(());
         })
     });
 }
@@ -245,14 +218,10 @@ where
                 // Create TX
                 let tx = db.tx_mut().expect("tx");
                 let mut crsr = tx.cursor_dup_write::<T>().expect("cursor");
-
+                for (k, _, v, _) in input {
-                black_box({
+                    crsr.append_dup(k, v).expect("submit");
-                    for (k, _, v, _) in input {
+                }
-                        crsr.append_dup(k, v).expect("submit");
+                tx.inner.commit().unwrap()
-                    }
-
-                    tx.inner.commit().unwrap()
-                });
             },
         )
     });
@@ -268,12 +237,10 @@ where
             |(input, db)| {
                 // Create TX
                 let tx = db.tx_mut().expect("tx");
-
                 for index in RANDOM_INDEXES {
                     let (k, _, v, _) = input.get(index).unwrap().clone();
                     tx.put::<T>(k, v).unwrap();
                 }
-
                 tx.inner.commit().unwrap();
             },
         )
@@ -286,14 +253,11 @@ where
             // Create TX
             let tx = db.tx().expect("tx");
 
-            {
+            let mut cursor = tx.cursor_dup_read::<T>().expect("cursor");
-                let mut cursor = tx.cursor_dup_read::<T>().expect("cursor");
+            let walker = cursor.walk_dup(None, Some(T::SubKey::default())).unwrap();
-                let walker = cursor.walk_dup(None, Some(T::SubKey::default())).unwrap();
+            for element in walker {
-                for element in walker {
+                element.unwrap();
-                    element.unwrap();
+            }
-                }
-            };
-            black_box(());
         })
     });
 

crates/storage/db/benches/hash_keys.rs

@@ -3,7 +3,7 @@
 use std::{collections::HashSet, path::Path, sync::Arc};
 
 use criterion::{
-    black_box, criterion_group, criterion_main, measurement::WallTime, BenchmarkGroup, Criterion,
+    criterion_group, criterion_main, measurement::WallTime, BenchmarkGroup, Criterion,
 };
 use pprof::criterion::{Output, PProfProfiler};
 use proptest::{
@@ -20,6 +20,7 @@ use reth_db_api::{
     transaction::DbTxMut,
 };
 use reth_fs_util as fs;
+use std::hint::black_box;
 
 mod utils;
 use utils::*;
@@ -46,6 +47,12 @@ pub fn hash_keys(c: &mut Criterion) {
     group.sample_size(10);
 
     for size in [10_000, 100_000, 1_000_000] {
+        // Too slow.
+        #[allow(unexpected_cfgs)]
+        if cfg!(codspeed) && size > 10_000 {
+            continue;
+        }
+
         measure_table_insertion::<TransactionHashNumbers>(&mut group, size);
     }
 }

crates/storage/db/benches/iai.rs

@@ -1,102 +0,0 @@
-#![allow(missing_docs, non_snake_case, unreachable_pub)]
-
-use iai_callgrind::{
-    library_benchmark, library_benchmark_group, LibraryBenchmarkConfig, RegressionConfig,
-};
-use paste::paste;
-use reth_db_api::table::{Compress, Decode, Decompress, Encode, Table};
-
-mod utils;
-use utils::*;
-
-macro_rules! impl_iai_callgrind_inner {
-    (
-        $(($name:ident, $group_name:ident, $mod:ident, $compress:ident, $decompress:ident, $encode:ident, $decode:ident, $seqread:ident, $randread:ident, $seqwrite:ident, $randwrite:ident))+
-    ) => {
-        use std::hint::black_box;
-        $(
-                #[library_benchmark]
-                pub fn $compress() {
-                    for (_, _, v, _) in black_box(load_vectors::<reth_db::tables::$name>()) {
-                        black_box(v.compress());
-                    }
-                }
-
-                #[library_benchmark]
-                pub fn $decompress() {
-                    for (_, _, _, comp) in black_box(load_vectors::<reth_db::tables::$name>()) {
-                        let _ = black_box(<reth_db::tables::$name as Table>::Value::decompress(&comp));
-                    }
-                }
-
-                #[library_benchmark]
-                pub fn $encode() {
-                    for (k, _, _, _) in black_box(load_vectors::<reth_db::tables::$name>()) {
-                        black_box(k.encode());
-                    }
-                }
-
-                #[library_benchmark]
-                pub fn $decode() {
-                    for (_, enc, _, _) in black_box(load_vectors::<reth_db::tables::$name>()) {
-                        let _ = black_box(<reth_db::tables::$name as Table>::Key::decode(&enc));
-                    }
-                }
-
-                #[allow(dead_code)]
-                pub const fn $seqread() {}
-
-                #[allow(dead_code)]
-                pub const fn $randread() {}
-
-                #[allow(dead_code)]
-                pub const fn $seqwrite() {}
-
-                #[allow(dead_code)]
-                pub const fn $randwrite() {}
-
-
-                library_benchmark_group!(
-                    name = $group_name;
-                    config = LibraryBenchmarkConfig::default()
-                        .regression(
-                            RegressionConfig::default().fail_fast(false)
-                        );
-                    benchmarks =
-                        $compress,
-                        $decompress,
-                        $encode,
-                        $decode,
-                );
-        )+
-
-        iai_callgrind::main!(
-                config = LibraryBenchmarkConfig::default();
-                library_benchmark_groups = $($group_name),+);
-    };
-}
-
-macro_rules! impl_iai_callgrind {
-    ($($name:ident),+) => {
-            paste! {
-                impl_iai_callgrind_inner!(
-                    $(
-                        ( $name, [<$name _group>],[<$name _mod>], [<$name _ValueCompress>], [<$name _ValueDecompress>], [<$name _ValueEncode>], [<$name _ValueDecode>], [<$name _SeqRead>], [<$name _RandomRead>], [<$name _SeqWrite>], [<$name _RandomWrite>])
-                    )+
-                );
-            }
-    };
-}
-
-impl_iai_callgrind!(
-    CanonicalHeaders,
-    HeaderTerminalDifficulties,
-    HeaderNumbers,
-    Headers,
-    BlockBodyIndices,
-    BlockOmmers,
-    TransactionHashNumbers,
-    Transactions,
-    PlainStorageState,
-    PlainAccountState
-);

crates/storage/db/src/lib.rs

@@ -60,15 +60,15 @@ pub mod test_utils {
     use tempfile::TempDir;
 
     /// Error during database open
-    pub const ERROR_DB_OPEN: &str = "Not able to open the database file.";
+    pub const ERROR_DB_OPEN: &str = "could not open the database file";
     /// Error during database creation
-    pub const ERROR_DB_CREATION: &str = "Not able to create the database file.";
+    pub const ERROR_DB_CREATION: &str = "could not create the database file";
     /// Error during database creation
-    pub const ERROR_STATIC_FILES_CREATION: &str = "Not able to create the static file path.";
+    pub const ERROR_STATIC_FILES_CREATION: &str = "could not create the static file path";
     /// Error during table creation
-    pub const ERROR_TABLE_CREATION: &str = "Not able to create tables in the database.";
+    pub const ERROR_TABLE_CREATION: &str = "could not create tables in the database";
     /// Error during tempdir creation
-    pub const ERROR_TEMPDIR: &str = "Not able to create a temporary directory.";
+    pub const ERROR_TEMPDIR: &str = "could not create a temporary directory";
 
     /// A database will delete the db dir when dropped.
     pub struct TempDatabase<DB> {

crates/storage/libmdbx-rs/benches/cursor.rs

@@ -1,10 +1,10 @@
 #![allow(missing_docs)]
 mod utils;
 
-use criterion::{black_box, criterion_group, criterion_main, Criterion};
+use criterion::{criterion_group, criterion_main, Criterion};
 use pprof::criterion::{Output, PProfProfiler};
 use reth_libmdbx::{ffi::*, *};
-use std::ptr;
+use std::{hint::black_box, ptr};
 use utils::*;
 
 /// Benchmark of iterator sequential read performance.

crates/storage/libmdbx-rs/benches/transaction.rs

@@ -1,11 +1,11 @@
 #![allow(missing_docs, unreachable_pub)]
 mod utils;
 
-use criterion::{black_box, criterion_group, criterion_main, Criterion};
+use criterion::{criterion_group, criterion_main, Criterion};
 use rand::{prelude::SliceRandom, SeedableRng};
 use rand_xorshift::XorShiftRng;
 use reth_libmdbx::{ffi::*, ObjectLength, WriteFlags};
-use std::ptr;
+use std::{hint::black_box, ptr};
 use utils::*;
 
 fn bench_get_rand(c: &mut Criterion) {

crates/transaction-pool/Cargo.toml

@@ -76,45 +76,45 @@ serde_json.workspace = true
 [features]
 default = ["serde"]
 serde = [
-	"dep:serde",
+    "dep:serde",
-	"reth-execution-types/serde",
+    "reth-execution-types/serde",
-	"reth-eth-wire-types/serde",
+    "reth-eth-wire-types/serde",
-	"reth-provider/serde",
+    "reth-provider/serde",
-	"alloy-consensus/serde",
+    "alloy-consensus/serde",
-	"alloy-eips/serde",
+    "alloy-eips/serde",
-	"alloy-primitives/serde",
+    "alloy-primitives/serde",
-	"bitflags/serde",
+    "bitflags/serde",
-	"parking_lot/serde",
+    "parking_lot/serde",
-	"rand?/serde",
+    "rand?/serde",
-	"smallvec/serde",
+    "smallvec/serde",
-	"reth-primitives-traits/serde",
+    "reth-primitives-traits/serde",
-	"revm-interpreter/serde",
+    "revm-interpreter/serde",
-	"revm-primitives/serde"
+    "revm-primitives/serde",
 ]
 test-utils = [
-	"rand",
+    "rand",
-	"paste",
+    "paste",
-	"serde",
+    "serde",
-	"reth-chain-state/test-utils",
+    "reth-chain-state/test-utils",
-	"reth-chainspec/test-utils",
+    "reth-chainspec/test-utils",
-	"reth-primitives/test-utils",
+    "reth-primitives/test-utils",
-	"reth-provider/test-utils",
+    "reth-provider/test-utils",
-	"reth-primitives-traits/test-utils",
+    "reth-primitives-traits/test-utils",
 ]
 arbitrary = [
-	"proptest",
+    "proptest",
-	"reth-primitives/arbitrary",
+    "reth-primitives/arbitrary",
-	"proptest-arbitrary-interop",
+    "proptest-arbitrary-interop",
-	"reth-chainspec/arbitrary",
+    "reth-chainspec/arbitrary",
-	"reth-eth-wire-types/arbitrary",
+    "reth-eth-wire-types/arbitrary",
-	"alloy-consensus/arbitrary",
+    "alloy-consensus/arbitrary",
-	"alloy-eips/arbitrary",
+    "alloy-eips/arbitrary",
-	"alloy-primitives/arbitrary",
+    "alloy-primitives/arbitrary",
-	"bitflags/arbitrary",
+    "bitflags/arbitrary",
-	"reth-primitives-traits/arbitrary",
+    "reth-primitives-traits/arbitrary",
-	"smallvec/arbitrary",
+    "smallvec/arbitrary",
-	"revm-interpreter/arbitrary",
+    "revm-interpreter/arbitrary",
-	"revm-primitives/arbitrary"
+    "revm-primitives/arbitrary",
 ]
 
 [[bench]]

crates/transaction-pool/benches/priority.rs

@@ -1,9 +1,10 @@
 #![allow(missing_docs)]
 use criterion::{
-    black_box, criterion_group, criterion_main, measurement::WallTime, BenchmarkGroup, Criterion,
+    criterion_group, criterion_main, measurement::WallTime, BenchmarkGroup, Criterion,
 };
 use proptest::{prelude::*, strategy::ValueTree, test_runner::TestRunner};
 use reth_transaction_pool::{blob_tx_priority, fee_delta};
+use std::hint::black_box;
 
 fn generate_test_data_fee_delta() -> (u128, u128) {
     let config = ProptestConfig::default();

crates/transaction-pool/benches/reorder.rs

@@ -75,19 +75,17 @@ fn txpool_reordering_bench<T: BenchTxPool>(
     );
     group.bench_function(group_id, |b| {
         b.iter_with_setup(setup, |(mut txpool, new_txs)| {
-            {
+            // Reorder with new base fee
-                // Reorder with new base fee
+            let bigger_base_fee = base_fee.saturating_add(10);
-                let bigger_base_fee = base_fee.saturating_add(10);
+            txpool.reorder(bigger_base_fee);
-                txpool.reorder(bigger_base_fee);
 
-                // Reorder with new base fee after adding transactions.
+            // Reorder with new base fee after adding transactions.
-                for new_tx in new_txs {
+            for new_tx in new_txs {
-                    txpool.add_transaction(new_tx);
+                txpool.add_transaction(new_tx);
-                }
+            }
-                let smaller_base_fee = base_fee.saturating_sub(10);
+            let smaller_base_fee = base_fee.saturating_sub(10);
-                txpool.reorder(smaller_base_fee)
+            txpool.reorder(smaller_base_fee);
-            };
+            txpool
-            std::hint::black_box(());
         });
     });
 }

crates/transaction-pool/benches/truncate.rs

@@ -161,7 +161,7 @@ fn truncate_pending(
     group.bench_function(group_id, |b| {
         b.iter_with_setup(setup, |mut txpool| {
             txpool.truncate_pool(SubPoolLimit::default());
-            std::hint::black_box(());
+            txpool
         });
     });
 }
@@ -195,7 +195,7 @@ fn truncate_queued(
     group.bench_function(group_id, |b| {
         b.iter_with_setup(setup, |mut txpool| {
             txpool.truncate_pool(SubPoolLimit::default());
-            std::hint::black_box(());
+            txpool
         });
     });
 }
@@ -229,7 +229,7 @@ fn truncate_basefee(
     group.bench_function(group_id, |b| {
         b.iter_with_setup(setup, |mut txpool| {
             txpool.truncate_pool(SubPoolLimit::default());
-            std::hint::black_box(());
+            txpool
         });
     });
 }

crates/trie/common/benches/prefix_set.rs

@@ -1,6 +1,6 @@
 #![allow(missing_docs, unreachable_pub)]
 use criterion::{
-    black_box, criterion_group, criterion_main, measurement::WallTime, BenchmarkGroup, Criterion,
+    criterion_group, criterion_main, measurement::WallTime, BenchmarkGroup, Criterion,
 };
 use proptest::{
     prelude::*,
@@ -11,7 +11,7 @@ use reth_trie_common::{
     prefix_set::{PrefixSet, PrefixSetMut},
     Nibbles,
 };
-use std::collections::BTreeSet;
+use std::{collections::BTreeSet, hint::black_box};
 
 /// Abstraction for aggregating nibbles and freezing it to a type
 /// that can be later used for benching.
@@ -48,6 +48,12 @@ pub fn prefix_set_lookups(c: &mut Criterion) {
     let mut group = c.benchmark_group("Prefix Set Lookups");
 
     for size in [10, 100, 1_000, 10_000] {
+        // Too slow.
+        #[allow(unexpected_cfgs)]
+        if cfg!(codspeed) && size > 1_000 {
+            continue;
+        }
+
         let test_data = generate_test_data(size);
 
         use implementations::*;

crates/trie/parallel/Cargo.toml

@@ -54,6 +54,14 @@ proptest-arbitrary-interop.workspace = true
 [features]
 default = ["metrics"]
 metrics = ["reth-metrics", "dep:metrics", "reth-trie/metrics"]
+test-utils = [
+    "reth-trie/test-utils",
+    "reth-trie-common/test-utils",
+    "reth-db/test-utils",
+    "reth-primitives/test-utils",
+    "reth-provider/test-utils",
+    "reth-trie-db/test-utils",
+]
 
 [[bench]]
 name = "root"

crates/trie/parallel/benches/root.rs

@@ -20,6 +20,12 @@ pub fn calculate_state_root(c: &mut Criterion) {
     group.sample_size(20);
 
     for size in [1_000, 3_000, 5_000, 10_000] {
+        // Too slow.
+        #[allow(unexpected_cfgs)]
+        if cfg!(codspeed) && size > 3_000 {
+            continue;
+        }
+
         let (db_state, updated_state) = generate_test_data(size);
         let provider_factory = create_test_provider_factory();
         {

crates/trie/sparse/Cargo.toml

@@ -41,6 +41,19 @@ proptest-arbitrary-interop.workspace = true
 proptest.workspace = true
 rand.workspace = true
 
+[features]
+test-utils = [
+    "reth-primitives-traits/test-utils",
+    "reth-trie/test-utils",
+    "reth-trie-common/test-utils",
+]
+arbitrary = [
+    "reth-primitives-traits/arbitrary",
+    "reth-trie-common/arbitrary",
+    "alloy-primitives/arbitrary",
+    "smallvec/arbitrary",
+]
+
 [[bench]]
 name = "root"
 harness = false

crates/trie/sparse/benches/rlp_node.rs

@@ -1,6 +1,4 @@
-#![allow(missing_docs, unreachable_pub)]
+#![allow(missing_docs)]
-
-use std::time::{Duration, Instant};
 
 use alloy_primitives::{B256, U256};
 use criterion::{criterion_group, criterion_main, Criterion};
@@ -11,7 +9,7 @@ use reth_testing_utils::generators;
 use reth_trie::Nibbles;
 use reth_trie_sparse::RevealedSparseTrie;
 
-pub fn update_rlp_node_level(c: &mut Criterion) {
+fn update_rlp_node_level(c: &mut Criterion) {
     let mut rng = generators::rng();
 
     let mut group = c.benchmark_group("update rlp node level");
@@ -53,20 +51,11 @@ pub fn update_rlp_node_level(c: &mut Criterion) {
                 group.bench_function(
                     format!("size {size} | updated {updated_leaves}% | depth {depth}"),
                     |b| {
-                        // Use `iter_custom` to avoid measuring clones and drops
+                        b.iter_batched_ref(
-                        b.iter_custom(|iters| {
+                            || sparse.clone(),
-                            let mut elapsed = Duration::ZERO;
+                            |cloned| cloned.update_rlp_node_level(depth),
-
+                            criterion::BatchSize::PerIteration,
-                            let mut cloned = sparse.clone();
+                        )
-                            for _ in 0..iters {
-                                let start = Instant::now();
-                                cloned.update_rlp_node_level(depth);
-                                elapsed += start.elapsed();
-                                cloned = sparse.clone();
-                            }
-
-                            elapsed
-                        })
                     },
                 );
             }

crates/trie/sparse/benches/root.rs

@@ -1,4 +1,4 @@
-#![allow(missing_docs, unreachable_pub)]
+#![allow(missing_docs)]
 
 use alloy_primitives::{map::B256HashMap, B256, U256};
 use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion};
@@ -15,11 +15,17 @@ use reth_trie::{
 use reth_trie_common::{HashBuilder, Nibbles};
 use reth_trie_sparse::SparseTrie;
 
-pub fn calculate_root_from_leaves(c: &mut Criterion) {
+fn calculate_root_from_leaves(c: &mut Criterion) {
     let mut group = c.benchmark_group("calculate root from leaves");
     group.sample_size(20);
 
     for size in [1_000, 5_000, 10_000, 100_000] {
+        // Too slow.
+        #[allow(unexpected_cfgs)]
+        if cfg!(codspeed) && size > 5_000 {
+            continue;
+        }
+
         let state = generate_test_data(size);
 
         // hash builder
@@ -29,6 +35,7 @@ pub fn calculate_root_from_leaves(c: &mut Criterion) {
                     hb.add_leaf(Nibbles::unpack(key), &alloy_rlp::encode_fixed_size(value));
                 }
                 hb.root();
+                hb
             })
         });
 
@@ -44,19 +51,32 @@ pub fn calculate_root_from_leaves(c: &mut Criterion) {
                         .unwrap();
                 }
                 sparse.root().unwrap();
+                sparse
             })
         });
     }
 }
 
-pub fn calculate_root_from_leaves_repeated(c: &mut Criterion) {
+fn calculate_root_from_leaves_repeated(c: &mut Criterion) {
     let mut group = c.benchmark_group("calculate root from leaves repeated");
     group.sample_size(20);
 
     for init_size in [1_000, 10_000, 100_000] {
+        // Too slow.
+        #[allow(unexpected_cfgs)]
+        if cfg!(codspeed) && init_size > 10_000 {
+            continue;
+        }
+
         let init_state = generate_test_data(init_size);
 
         for update_size in [100, 1_000, 5_000, 10_000] {
+            // Too slow.
+            #[allow(unexpected_cfgs)]
+            if cfg!(codspeed) && update_size > 1_000 {
+                continue;
+            }
+
             for num_updates in [1, 3, 5, 10] {
                 let updates =
                     (0..num_updates).map(|_| generate_test_data(update_size)).collect::<Vec<_>>();

crates/trie/trie/Cargo.toml

@@ -61,19 +61,20 @@ criterion.workspace = true
 [features]
 metrics = ["reth-metrics", "dep:metrics"]
 serde = [
-	"alloy-primitives/serde",
+    "alloy-primitives/serde",
-	"alloy-consensus/serde",
+    "alloy-consensus/serde",
-	"alloy-trie/serde",
+    "alloy-trie/serde",
     "alloy-eips/serde",
-	"revm/serde",
+    "revm/serde",
-	"reth-trie-common/serde"
+    "reth-trie-common/serde",
 ]
 test-utils = [
-	"triehash",
+    "triehash",
-	"revm/test-utils",
+    "revm/test-utils",
-	"reth-primitives/test-utils",
+    "reth-primitives/test-utils",
-	"reth-trie-common/test-utils",
+    "reth-trie-common/test-utils",
-	"reth-stages-types/test-utils"
+    "reth-trie-sparse/test-utils",
+    "reth-stages-types/test-utils",
 ]
 
 [[bench]]

crates/trie/trie/benches/hash_post_state.rs

@@ -10,6 +10,12 @@ pub fn hash_post_state(c: &mut Criterion) {
     group.sample_size(20);
 
     for size in [100, 1_000, 3_000, 5_000, 10_000] {
+        // Too slow.
+        #[allow(unexpected_cfgs)]
+        if cfg!(codspeed) && size > 1_000 {
+            continue;
+        }
+
         let state = generate_test_data(size);
 
         // sequence

crates/trie/trie/benches/trie_root.rs

@@ -1,10 +1,11 @@
 #![allow(missing_docs, unreachable_pub)]
 use alloy_primitives::B256;
-use criterion::{black_box, criterion_group, criterion_main, Criterion};
+use criterion::{criterion_group, criterion_main, Criterion};
 use proptest::{prelude::*, strategy::ValueTree, test_runner::TestRunner};
 use proptest_arbitrary_interop::arb;
 use reth_primitives::{Receipt, ReceiptWithBloom};
 use reth_trie::triehash::KeccakHasher;
+use std::hint::black_box;
 
 /// Benchmarks different implementations of the root calculation.
 pub fn trie_root_benchmark(c: &mut Criterion) {

docs/design/database.md

@@ -9,7 +9,7 @@
 
 - We want Reth's serialized format to be able to trade off read/write speed for size, depending on who the user is.
 - To achieve that, we created the [Encode/Decode/Compress/Decompress traits](https://github.com/paradigmxyz/reth/blob/0d9b9a392d4196793736522f3fc2ac804991b45d/crates/interfaces/src/db/table.rs#L9-L36) to make the (de)serialization of database `Table::Key` and `Table::Values` generic.
-  - This allows for [out-of-the-box benchmarking](https://github.com/paradigmxyz/reth/blob/0d9b9a392d4196793736522f3fc2ac804991b45d/crates/db/benches/encoding_iai.rs#L5) (using [Criterion](https://github.com/bheisler/criterion.rs) and [Iai](https://github.com/bheisler/iai))
+  - This allows for [out-of-the-box benchmarking](https://github.com/paradigmxyz/reth/blob/0d9b9a392d4196793736522f3fc2ac804991b45d/crates/db/benches/encoding_iai.rs#L5) (using [Criterion](https://github.com/bheisler/criterion.rs))
   - It also enables [out-of-the-box fuzzing](https://github.com/paradigmxyz/reth/blob/0d9b9a392d4196793736522f3fc2ac804991b45d/crates/interfaces/src/db/codecs/fuzz/mod.rs) using [trailofbits/test-fuzz](https://github.com/trailofbits/test-fuzz).
 - We implemented that trait for the following encoding formats:
   - [Ethereum-specific Compact Encoding](https://github.com/paradigmxyz/reth/blob/0d9b9a392d4196793736522f3fc2ac804991b45d/crates/codecs/derive/src/compact/mod.rs): A lot of Ethereum datatypes have unnecessary zeros when serialized, or optional (e.g. on empty hashes) which would be nice not to pay in storage costs.