Files
solana-streamer/src/streaming/common/event_processor.rs
T
ysq 74781e5cbe perf: implement SIMD-accelerated event processing and optimize streaming performance
- Add SIMD utilities for fast byte array comparison and discriminator matching
- Optimize event processor with batch processing and memory pool
- Refactor global state management with concurrent data structures
- Remove deprecated batch processing module
- Enhance metrics collection with reduced overhead
- Improve parser efficiency across all protocol implementations
- Add performance benchmarking dependencies (criterion, wide)
- Update documentation and examples for new architecture

Performance improvements:
- SIMD-accelerated byte operations for instruction parsing
- Concurrent HashMap (DashMap) for better multi-threading
- Optimized memory allocation patterns
- Reduced lock contention in event processing pipeline

Breaking changes: Removed batch.rs module, updated parser interfaces
2025-08-31 22:18:18 +08:00

435 lines
18 KiB
Rust

use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
use std::sync::Arc;
use std::time::Instant;
use crossbeam_queue::SegQueue;
use solana_sdk::pubkey::Pubkey;
use crate::common::AnyResult;
use crate::streaming::common::BackpressureStrategy;
use crate::streaming::common::{
MetricsEventType, MetricsManager, StreamClientConfig as ClientConfig,
};
use crate::streaming::event_parser::common::filter::EventTypeFilter;
use crate::streaming::event_parser::core::account_event_parser::AccountEventParser;
use crate::streaming::event_parser::core::common_event_parser::CommonEventParser;
use crate::streaming::event_parser::core::traits::get_high_perf_clock;
use crate::streaming::event_parser::EventParser;
use crate::streaming::event_parser::{
core::traits::UnifiedEvent, protocols::mutil::parser::MutilEventParser, Protocol,
};
use crate::streaming::grpc::{BackpressureConfig, EventPretty};
use crate::streaming::shred::TransactionWithSlot;
use once_cell::sync::OnceCell;
/// High-performance Event processor using SegQueue for all strategies
pub struct EventProcessor {
pub(crate) metrics_manager: MetricsManager,
pub(crate) config: ClientConfig,
pub(crate) parser_cache: OnceCell<Arc<dyn EventParser>>,
pub(crate) protocols: Vec<Protocol>,
pub(crate) event_type_filter: Option<EventTypeFilter>,
pub(crate) callback: Option<Arc<dyn Fn(Box<dyn UnifiedEvent>) + Send + Sync>>,
pub(crate) backpressure_config: BackpressureConfig,
/// High-performance lockfree queue for gRPC events
pub(crate) grpc_queue: Arc<SegQueue<(EventPretty, Option<Pubkey>)>>,
/// High-performance lockfree queue for shred events
pub(crate) shred_queue: Arc<SegQueue<(TransactionWithSlot, Option<Pubkey>)>>,
/// Fast O(1) counter for Drop strategy (avoids expensive SegQueue::len())
pub(crate) grpc_pending_count: Arc<AtomicUsize>,
pub(crate) shred_pending_count: Arc<AtomicUsize>,
/// Processing thread control
pub(crate) processing_shutdown: Arc<AtomicBool>,
}
impl EventProcessor {
/// Create a new high-performance event processor
pub fn new(metrics_manager: MetricsManager, config: ClientConfig) -> Self {
let backpressure_config = config.backpressure.clone();
let grpc_queue = Arc::new(SegQueue::new());
let shred_queue = Arc::new(SegQueue::new());
let grpc_pending_count = Arc::new(AtomicUsize::new(0));
let shred_pending_count = Arc::new(AtomicUsize::new(0));
let processing_shutdown = Arc::new(AtomicBool::new(false));
Self {
metrics_manager,
config,
parser_cache: OnceCell::new(),
protocols: vec![],
event_type_filter: None,
backpressure_config,
callback: None,
grpc_queue,
shred_queue,
grpc_pending_count,
shred_pending_count,
processing_shutdown,
}
}
pub fn set_protocols_and_event_type_filter(
&mut self,
protocols: Vec<Protocol>,
event_type_filter: Option<EventTypeFilter>,
backpressure_config: BackpressureConfig,
callback: Option<Arc<dyn Fn(Box<dyn UnifiedEvent>) + Send + Sync>>,
) {
self.protocols = protocols;
self.event_type_filter = event_type_filter;
// Check if Block processing thread should be started (before moving backpressure_config)
let should_start_block_processing = true;
// matches!(backpressure_config.strategy, BackpressureStrategy::Block);
self.backpressure_config = backpressure_config;
self.callback = callback;
// Use stored values to initialize parser_cache
let protocols_ref = &self.protocols;
let event_type_filter_ref = self.event_type_filter.as_ref();
self.parser_cache.get_or_init(|| {
Arc::new(MutilEventParser::new(protocols_ref.clone(), event_type_filter_ref.cloned()))
});
// Start Block processing thread if using Block strategy
if should_start_block_processing {
self.start_block_processing_thread();
}
}
pub fn get_parser(&self) -> Arc<dyn EventParser> {
self.parser_cache.get().unwrap().clone()
}
/// Create adapter callback
fn create_adapter_callback(&self) -> Arc<dyn Fn(Box<dyn UnifiedEvent>) + Send + Sync> {
let callback = self.callback.clone().unwrap();
let metrics_manager = self.metrics_manager.clone();
Arc::new(move |event: Box<dyn UnifiedEvent>| {
let processing_time_us = event.program_handle_time_consuming_us() as f64;
callback(event);
metrics_manager.update_metrics(MetricsEventType::Transaction, 1, processing_time_us);
})
}
pub async fn process_grpc_event_transaction_with_metrics(
&self,
event_pretty: EventPretty,
bot_wallet: Option<Pubkey>,
) -> AnyResult<()> {
self.apply_backpressure_control(event_pretty, bot_wallet).await
}
/// Apply backpressure control strategy
async fn apply_backpressure_control(
&self,
event_pretty: EventPretty,
bot_wallet: Option<Pubkey>,
) -> AnyResult<()> {
match self.backpressure_config.strategy {
BackpressureStrategy::Block => {
// Block strategy: async wait if queue is full (backpressure control)
loop {
let current_pending = self.grpc_pending_count.load(Ordering::Relaxed);
if current_pending < self.backpressure_config.permits {
self.grpc_queue.push((event_pretty, bot_wallet));
self.grpc_pending_count.fetch_add(1, Ordering::Relaxed);
break;
}
// Async yield to avoid blocking gRPC data source
tokio::task::yield_now().await;
}
Ok(())
}
BackpressureStrategy::Drop => {
// Drop strategy: Use O(1) atomic counter instead of expensive O(n) len()
// If pending count >= permits, DROP the event immediately
let current_pending = self.grpc_pending_count.load(Ordering::Relaxed);
if current_pending >= self.backpressure_config.permits {
self.metrics_manager.increment_dropped_events();
Ok(())
} else {
self.grpc_pending_count.fetch_add(1, Ordering::Relaxed);
let processor = self.clone();
tokio::spawn(async move {
match processor
.process_grpc_event_transaction(event_pretty, bot_wallet)
.await
{
Ok(_) => {
processor.grpc_pending_count.fetch_sub(1, Ordering::Relaxed);
}
Err(e) => {
log::error!("Error in async gRPC processing: {}", e);
}
}
});
Ok(())
}
}
}
}
async fn process_grpc_event_transaction(
&self,
event_pretty: EventPretty,
bot_wallet: Option<Pubkey>,
) -> AnyResult<()> {
if self.callback.is_none() {
return Ok(());
}
match event_pretty {
EventPretty::Account(account_pretty) => {
self.metrics_manager.add_account_process_count();
let account_event = AccountEventParser::parse_account_event(
&self.protocols,
account_pretty,
self.event_type_filter.as_ref(),
);
if let Some(event) = account_event {
let processing_time_us = event.program_handle_time_consuming_us() as f64;
self.invoke_callback(event);
self.update_metrics(MetricsEventType::Account, 1, processing_time_us);
}
}
EventPretty::Transaction(transaction_pretty) => {
self.metrics_manager.add_tx_process_count();
let slot = transaction_pretty.slot;
let signature = transaction_pretty.signature;
let tx = transaction_pretty.tx;
let block_time = transaction_pretty.block_time;
let program_received_time_us = transaction_pretty.program_received_time_us;
let transaction_index = transaction_pretty.transaction_index;
// Use cache to get parser
let parser = self.get_parser();
let adapter_callback = self.create_adapter_callback();
parser
.parse_transaction_owned(
tx,
signature,
Some(slot),
block_time,
program_received_time_us,
bot_wallet,
transaction_index,
adapter_callback,
)
.await?;
}
EventPretty::BlockMeta(block_meta_pretty) => {
self.metrics_manager.add_block_meta_process_count();
let block_time_ms = block_meta_pretty
.block_time
.map(|ts| ts.seconds * 1000 + ts.nanos as i64 / 1_000_000)
.unwrap_or_else(|| chrono::Utc::now().timestamp_millis());
let block_meta_event = CommonEventParser::generate_block_meta_event(
block_meta_pretty.slot,
&block_meta_pretty.block_hash,
block_time_ms,
block_meta_pretty.program_received_time_us,
);
let processing_time_us = block_meta_event.program_handle_time_consuming_us() as f64;
self.invoke_callback(block_meta_event);
self.update_metrics(MetricsEventType::BlockMeta, 1, processing_time_us);
}
}
Ok(())
}
pub fn invoke_callback(&self, event: Box<dyn UnifiedEvent>) {
if let Some(callback) = self.callback.as_ref() {
callback(event);
}
}
/// Process a single transaction immediately
pub async fn process_shred_transaction_immediate(
&self,
transaction_with_slot: TransactionWithSlot,
bot_wallet: Option<Pubkey>,
) -> AnyResult<()> {
self.process_shred_transaction(transaction_with_slot, bot_wallet).await
}
/// Process shred transaction with backpressure control and performance monitoring
pub async fn process_shred_transaction_with_metrics(
&self,
transaction_with_slot: TransactionWithSlot,
bot_wallet: Option<Pubkey>,
) -> AnyResult<()> {
// Backpressure control logic
self.apply_shred_backpressure_control(transaction_with_slot, bot_wallet).await
}
/// Apply shred backpressure control strategy
async fn apply_shred_backpressure_control(
&self,
transaction_with_slot: TransactionWithSlot,
bot_wallet: Option<Pubkey>,
) -> AnyResult<()> {
match self.backpressure_config.strategy {
BackpressureStrategy::Block => {
// Block strategy: async wait if queue is full (backpressure control)
loop {
let current_pending = self.shred_pending_count.load(Ordering::Relaxed);
if current_pending < self.backpressure_config.permits {
self.shred_queue.push((transaction_with_slot, bot_wallet));
self.shred_pending_count.fetch_add(1, Ordering::Relaxed);
break;
}
// Async yield to avoid blocking shred data source
tokio::task::yield_now().await;
}
Ok(())
}
BackpressureStrategy::Drop => {
// Drop strategy: Use O(1) atomic counter instead of expensive O(n) len()
let current_pending = self.shred_pending_count.load(Ordering::Relaxed);
if current_pending >= self.backpressure_config.permits {
self.metrics_manager.increment_dropped_events();
Ok(())
} else {
self.shred_pending_count.fetch_add(1, Ordering::Relaxed);
let processor = self.clone();
tokio::spawn(async move {
match processor
.process_shred_transaction(transaction_with_slot, bot_wallet)
.await
{
Ok(_) => {
processor.shred_pending_count.fetch_sub(1, Ordering::Relaxed);
}
Err(e) => {
log::error!("Error in async shred processing: {}", e);
}
}
});
Ok(())
}
}
}
}
pub async fn process_shred_transaction(
&self,
transaction_with_slot: TransactionWithSlot,
bot_wallet: Option<Pubkey>,
) -> AnyResult<()> {
if self.callback.is_none() {
return Ok(());
}
self.metrics_manager.add_tx_process_count();
let tx = transaction_with_slot.transaction;
let slot = transaction_with_slot.slot;
let signature = tx.signatures[0];
let program_received_time_us = transaction_with_slot.program_received_time_us;
// Use cache to get parser
let parser = self.get_parser();
let adapter_callback = self.create_adapter_callback();
parser
.parse_versioned_transaction_owned(
tx,
signature,
Some(slot),
None,
program_received_time_us,
bot_wallet,
None,
&[],
adapter_callback,
)
.await?;
Ok(())
}
fn update_metrics(&self, ty: MetricsEventType, count: u64, time_us: f64) {
self.metrics_manager.update_metrics(ty, count, time_us);
}
/// Start dedicated processing threads for all strategies
fn start_block_processing_thread(&self) {
// Reset shutdown flag
self.processing_shutdown.store(false, Ordering::Relaxed);
let grpc_queue = Arc::clone(&self.grpc_queue);
let shred_queue = Arc::clone(&self.shred_queue);
let grpc_pending_count = Arc::clone(&self.grpc_pending_count);
let shred_pending_count = Arc::clone(&self.shred_pending_count);
let shutdown_flag = Arc::clone(&self.processing_shutdown);
let shutdown_flag_clone = Arc::clone(&self.processing_shutdown);
let processor = self.clone();
let processor_clone = self.clone();
// 1. 专用线程 + 2. Busy-wait + 4. 无锁处理
std::thread::spawn(move || {
// 创建blocking runtime for async processing
let rt = tokio::runtime::Builder::new_current_thread().enable_all().build().unwrap();
while !shutdown_flag.load(Ordering::Relaxed) {
if let Some((event_pretty, bot_wallet)) = grpc_queue.pop() {
// Decrement pending counter when consuming from queue
grpc_pending_count.fetch_sub(1, Ordering::Relaxed);
// Process event in blocking runtime
if let Err(e) = rt.block_on(
processor.process_grpc_event_transaction(event_pretty, bot_wallet),
) {
println!("Error processing gRPC event: {}", e);
}
} else {
// 2. 优化忙等待: 使用轻量级休眠减少CPU占用
std::thread::yield_now();
}
}
});
// Shred处理也使用相同的低延迟优化
std::thread::spawn(move || {
let rt = tokio::runtime::Builder::new_current_thread().enable_all().build().unwrap();
while !shutdown_flag_clone.load(Ordering::Relaxed) {
if let Some((transaction_with_slot, bot_wallet)) = shred_queue.pop() {
// Decrement pending counter when consuming from queue
shred_pending_count.fetch_sub(1, Ordering::Relaxed);
// Process transaction in blocking runtime
if let Err(e) = rt.block_on(
processor_clone
.process_shred_transaction(transaction_with_slot, bot_wallet),
) {
log::error!("Error processing shred transaction: {}", e);
}
} else {
// 优化忙等待: 使用轻量级休眠减少CPU占用
std::thread::yield_now();
}
}
});
}
/// Stop processing threads
pub fn stop_processing(&self) {
self.processing_shutdown.store(true, Ordering::Relaxed);
}
}
// Implement Clone trait to support sharing between modules
impl Clone for EventProcessor {
fn clone(&self) -> Self {
Self {
metrics_manager: self.metrics_manager.clone(),
config: self.config.clone(),
parser_cache: self.parser_cache.clone(),
protocols: self.protocols.clone(),
event_type_filter: self.event_type_filter.clone(),
backpressure_config: self.backpressure_config.clone(),
callback: self.callback.clone(),
grpc_queue: self.grpc_queue.clone(),
shred_queue: self.shred_queue.clone(),
grpc_pending_count: self.grpc_pending_count.clone(),
shred_pending_count: self.shred_pending_count.clone(),
processing_shutdown: self.processing_shutdown.clone(),
}
}
}