optimize state cache

src/lib.rs

@@ -1,6 +1,5 @@
 use lazy_static::lazy_static;
-use revm::primitives::{Address, EVMError, EVMResultGeneric, ExecutionResult, U256};
-use revm::TransitionAccount;
+use revm::primitives::{Address, EVMError, EVMResultGeneric, EvmState, ExecutionResult, U256};
 use std::cmp::min;
 use std::fmt::{Display, Formatter};
 use std::sync::atomic::{AtomicUsize, Ordering};
@@ -13,13 +12,14 @@ mod storage;
 mod tx_dependency;
 
 lazy_static! {
-    static ref CPU_CORES: usize = thread::available_parallelism().map(|n| n.get()).unwrap_or(8);
+    static ref CPU_CORES: usize =
+        std::thread::available_parallelism().map(|n| n.get()).unwrap_or(8);
 }
 
 lazy_static! {
     static ref GREVM_RUNTIME: Runtime = Builder::new_multi_thread()
         // .worker_threads(1) // for debug
-        .worker_threads(thread::available_parallelism().map(|n| n.get() * 2).unwrap_or(8))
+        .worker_threads(std::thread::available_parallelism().map(|n| n.get() * 2).unwrap_or(8))
         .thread_name("grevm-tokio-runtime")
         .enable_all()
         .build()
@@ -75,7 +75,7 @@ pub struct ResultAndTransition {
     /// Status of execution
     pub result: Option<ExecutionResult>,
     /// State that got updated
-    pub transition: Vec<(Address, TransitionAccount)>,
+    pub transition: EvmState,
     /// Rewards to miner
     pub rewards: u128,
 }

src/partition.rs

@@ -156,7 +156,12 @@ where
                     if read_set.is_disjoint(&update_write_set) {
                         self.read_set.push(read_set);
                         self.write_set.push(write_set);
-                        evm.db_mut().temporary_commit_transition(&execute_state.transition);
+                        // All accounts should be present inside cache.
+                        for address in execute_state.transition.keys() {
+                            // FIXME(gravity): error handling
+                            let _ = evm.db_mut().load_cache_account(*address);
+                        }
+                        evm.db_mut().temporary_commit(&execute_state.transition);
                         self.execute_results.push(Ok(execute_state));
                         should_rerun = false;
                         self.metrics.reusable_tx_cnt += 1;
@@ -182,10 +187,10 @@ where
                             result_and_state.state.remove(&self.coinbase);
                         }
                         // temporary commit to cache_db, to make use the remaining txs can read the updated data
-                        let transition = evm.db_mut().temporary_commit(result_and_state.state);
+                        evm.db_mut().temporary_commit(&result_and_state.state);
                         self.execute_results.push(Ok(ResultAndTransition {
                             result: Some(result_and_state.result),
-                            transition,
+                            transition: result_and_state.state,
                             rewards: rewards.unwrap_or(0),
                         }));
                     }

src/scheduler.rs

@@ -9,20 +9,19 @@ use crate::hint::ParallelExecutionHints;
 use crate::partition::{
     OrderedVectorExt, PartitionExecutor, PreRoundContext, PreUnconfirmedContext,
 };
-use crate::storage::SchedulerDB;
+use crate::storage::{SchedulerDB, StateCache};
 use crate::tx_dependency::{DependentTxsVec, TxDependency};
 use crate::{
-    fork_join_util, GrevmError, LocationAndType, PartitionId, TxId, CPU_CORES, GREVM_RUNTIME,
-    MAX_NUM_ROUND,
+    fork_join_util, GrevmError, LocationAndType, TxId, CPU_CORES, GREVM_RUNTIME, MAX_NUM_ROUND,
 };
 
-use metrics::{counter, gauge, histogram};
+use metrics::{counter, gauge};
 use revm::db::states::bundle_state::BundleRetention;
 use revm::db::BundleState;
 use revm::primitives::{
     AccountInfo, Address, Bytecode, EVMError, Env, ExecutionResult, SpecId, TxEnv, B256, U256,
 };
-use revm::{CacheState, DatabaseRef, EvmBuilder};
+use revm::{DatabaseRef, EvmBuilder};
 
 pub struct ExecuteMetrics {
     /// Number of times parallel execution is called.
@@ -460,12 +459,20 @@ where
         // update and pruning tx dependencies
         self.update_and_pruning_dependency();
 
-        let partition_state: Vec<CacheState> = self
+        let partition_state: Vec<StateCache> = self
             .partition_executors
             .iter()
             .map(|executor| {
                 let mut executor = executor.write().unwrap();
-                std::mem::take(&mut executor.partition_db.cache)
+                let mut cache = std::mem::take(&mut executor.partition_db.cache);
+                if !executor.partition_db.miner_involved {
+                    // Remove miner account in partitions that only treat it as a reward account.
+                    // We have preload it into scheduler state cache, and will handle rewards
+                    // separately. Only when miner account participates in the transactions should
+                    // it be merged into the scheduler state cache.
+                    cache.accounts.remove(&self.coinbase);
+                }
+                cache
             })
             .collect();
 
@@ -475,14 +482,40 @@ where
         drop(span);
         let database = Arc::get_mut(&mut self.database).unwrap();
 
-        Self::merge_not_modified_state(&mut database.cache, partition_state);
+        if self.num_finality_txs == self.txs.len() {
+            // Merge partition state to scheduler directly if there are no conflicts in the entire
+            // partition.
+            let span = Span::enter_with_local_parent("directly merge accounts");
+            for state in partition_state {
+                database.cache.merge_accounts(state);
+            }
+
+            let mut rewards: u128 = 0;
+            for ctx in finality_txs.into_values() {
+                rewards += ctx.execute_state.rewards;
+                self.results.push(ctx.execute_state.result.unwrap());
+            }
+            database
+                .increment_balances(vec![(self.coinbase, rewards)])
+                .map_err(|err| GrevmError::EvmError(EVMError::Database(err)))?;
+
+            self.metrics.validate_time.increment(start.elapsed().as_nanos() as u64);
+            return Ok(());
+        }
+
+        // Must merge loaded state to database cache before commit changes
+        let span = Span::enter_with_local_parent("merge loaded state");
+        for state in partition_state {
+            database.cache.merge_loaded_state(state);
+        }
+        std::mem::drop(span);
 
         let span = Span::enter_with_local_parent("commit transition to scheduler db");
         let mut rewards: u128 = 0;
         for ctx in finality_txs.into_values() {
             rewards += ctx.execute_state.rewards;
             self.results.push(ctx.execute_state.result.unwrap());
-            database.commit_transition(ctx.execute_state.transition);
+            database.commit(&ctx.execute_state.transition);
         }
         // Each transaction updates three accounts: from, to, and coinbase.
         // If every tx updates the coinbase account, it will cause conflicts across all txs.
@@ -498,27 +531,6 @@ where
         Ok(())
     }
 
-    /// Merge not modified state from partition to scheduler. These data are just loaded from
-    /// database, so we can merge them to state as original value for next round.
-    #[fastrace::trace]
-    fn merge_not_modified_state(state: &mut CacheState, partition_state: Vec<CacheState>) {
-        for partition in partition_state {
-            // merge account state that is not modified
-            for (address, account) in partition.accounts {
-                if account.status.is_not_modified() && state.accounts.get(&address).is_none() {
-                    state.accounts.insert(address, account);
-                }
-            }
-
-            // merge contract code
-            for (hash, code) in partition.contracts {
-                if state.contracts.get(&hash).is_none() {
-                    state.contracts.insert(hash, code);
-                }
-            }
-        }
-    }
-
     #[fastrace::trace]
     fn execute_remaining_sequential(&mut self) -> Result<(), GrevmError<DB::Error>> {
         self.metrics.sequential_execute_calls.increment(1);
@@ -539,7 +551,7 @@ where
             }
             match evm.transact() {
                 Ok(result_and_state) => {
-                    evm.db_mut().commit(result_and_state.state);
+                    evm.db_mut().commit(&result_and_state.state);
                     self.results.push(result_and_state.result);
                 }
                 Err(err) => return Err(GrevmError::EvmError(err)),
@@ -562,13 +574,18 @@ where
             // MUST drop the `PartitionExecutor::scheduler_db` before get mut
             this.partition_executors.clear();
             let database = Arc::get_mut(&mut this.database).unwrap();
-            database.merge_transitions(BundleRetention::Reverts);
-            ExecuteOutput {
-                state: std::mem::take(&mut database.bundle_state),
-                results: std::mem::take(&mut this.results),
-            }
+            let state = database.create_bundle_state(BundleRetention::Reverts);
+            ExecuteOutput { state, results: std::mem::take(&mut this.results) }
         };
 
+        {
+            // Preload coinbase account
+            let db = Arc::get_mut(&mut self.database).unwrap();
+            let _ = db
+                .load_cache_account(self.coinbase)
+                .map_err(|err| GrevmError::EvmError(EVMError::Database(err)))?;
+        }
+
         if self.txs.len() < self.num_partitions && !force_parallel {
             self.execute_remaining_sequential()?;
             return Ok(build_execute_output(self));