Leader QoS service metrics (#21708)

* - qos_service metrics tagged with leader thread ids to separate gossip/tpu votes and transactions;
- qos_service metrics is reported with bank slot;
- replaced timer-based reporting with signal via channel; removed async report test as qos_service now lives within a thread

* - add tpu live packets (eg, not buffered packets) states to qos metrics reporting
This commit is contained in:
Tao Zhu 2021-12-22 15:39:59 -06:00 committed by GitHub
parent 2be139ca61
commit dd80a525ef
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 174 additions and 118 deletions

View File

@ -97,7 +97,7 @@ fn bench_consume_buffered(bencher: &mut Bencher) {
None::<Box<dyn Fn()>>,
&BankingStageStats::default(),
&recorder,
&Arc::new(QosService::new(Arc::new(RwLock::new(CostModel::default())))),
&QosService::new(Arc::new(RwLock::new(CostModel::default())), 1),
);
});

View File

@ -11,7 +11,7 @@ use {
solana_gossip::{cluster_info::ClusterInfo, contact_info::ContactInfo},
solana_ledger::blockstore_processor::TransactionStatusSender,
solana_measure::measure::Measure,
solana_metrics::{inc_new_counter_debug, inc_new_counter_info},
solana_metrics::inc_new_counter_info,
solana_perf::{
cuda_runtime::PinnedVec,
data_budget::DataBudget,
@ -330,7 +330,6 @@ impl BankingStage {
PacketHasher::default(),
)));
let data_budget = Arc::new(DataBudget::default());
let qos_service = Arc::new(QosService::new(cost_model));
// Many banks that process transactions in parallel.
assert!(num_threads >= NUM_VOTE_PROCESSING_THREADS + MIN_THREADS_BANKING);
let bank_thread_hdls: Vec<JoinHandle<()>> = (0..num_threads)
@ -355,7 +354,7 @@ impl BankingStage {
let gossip_vote_sender = gossip_vote_sender.clone();
let duplicates = duplicates.clone();
let data_budget = data_budget.clone();
let qos_service = qos_service.clone();
let cost_model = cost_model.clone();
Builder::new()
.name("solana-banking-stage-tx".to_string())
.spawn(move || {
@ -371,7 +370,7 @@ impl BankingStage {
gossip_vote_sender,
&duplicates,
&data_budget,
qos_service,
cost_model,
);
})
.unwrap()
@ -445,7 +444,7 @@ impl BankingStage {
test_fn: Option<impl Fn()>,
banking_stage_stats: &BankingStageStats,
recorder: &TransactionRecorder,
qos_service: &Arc<QosService>,
qos_service: &QosService,
) {
let mut rebuffered_packet_count = 0;
let mut new_tx_count = 0;
@ -502,6 +501,7 @@ impl BankingStage {
));
}
new_tx_count += processed;
// Out of the buffered packets just retried, collect any still unprocessed
// transactions in this batch for forwarding
rebuffered_packet_count += new_unprocessed_indexes.len();
@ -594,7 +594,7 @@ impl BankingStage {
banking_stage_stats: &BankingStageStats,
recorder: &TransactionRecorder,
data_budget: &DataBudget,
qos_service: &Arc<QosService>,
qos_service: &QosService,
) -> BufferedPacketsDecision {
let bank_start;
let (
@ -714,12 +714,13 @@ impl BankingStage {
gossip_vote_sender: ReplayVoteSender,
duplicates: &Arc<Mutex<(LruCache<u64, ()>, PacketHasher)>>,
data_budget: &DataBudget,
qos_service: Arc<QosService>,
cost_model: Arc<RwLock<CostModel>>,
) {
let recorder = poh_recorder.lock().unwrap().recorder();
let socket = UdpSocket::bind("0.0.0.0:0").unwrap();
let mut buffered_packet_batches = VecDeque::with_capacity(batch_limit);
let banking_stage_stats = BankingStageStats::new(id);
let qos_service = QosService::new(cost_model, id);
loop {
let my_pubkey = cluster_info.id();
while !buffered_packet_batches.is_empty() {
@ -968,7 +969,7 @@ impl BankingStage {
chunk_offset: usize,
transaction_status_sender: Option<TransactionStatusSender>,
gossip_vote_sender: &ReplayVoteSender,
qos_service: &Arc<QosService>,
qos_service: &QosService,
) -> (Result<usize, PohRecorderError>, Vec<usize>) {
let tx_costs = qos_service.compute_transaction_costs(txs.iter());
@ -999,6 +1000,9 @@ impl BankingStage {
drop(batch);
unlock_time.stop();
// reports qos service stats for this batch
qos_service.report_metrics(bank.clone());
debug!(
"bank: {} lock: {}us unlock: {}us txs_len: {}",
bank.slot(),
@ -1021,7 +1025,7 @@ impl BankingStage {
poh: &TransactionRecorder,
transaction_status_sender: Option<TransactionStatusSender>,
gossip_vote_sender: &ReplayVoteSender,
qos_service: &Arc<QosService>,
qos_service: &QosService,
) -> (usize, Vec<usize>) {
let mut chunk_start = 0;
let mut unprocessed_txs = vec![];
@ -1190,7 +1194,7 @@ impl BankingStage {
transaction_status_sender: Option<TransactionStatusSender>,
gossip_vote_sender: &ReplayVoteSender,
banking_stage_stats: &BankingStageStats,
qos_service: &Arc<QosService>,
qos_service: &QosService,
) -> (usize, usize, Vec<usize>) {
let mut packet_conversion_time = Measure::start("packet_conversion");
let (transactions, transaction_to_packet_indexes) = Self::transactions_from_packets(
@ -1330,7 +1334,7 @@ impl BankingStage {
banking_stage_stats: &BankingStageStats,
duplicates: &Arc<Mutex<(LruCache<u64, ()>, PacketHasher)>>,
recorder: &TransactionRecorder,
qos_service: &Arc<QosService>,
qos_service: &QosService,
) -> Result<(), RecvTimeoutError> {
let mut recv_time = Measure::start("process_packets_recv");
let packet_batches = verified_receiver.recv_timeout(recv_timeout)?;
@ -1345,7 +1349,6 @@ impl BankingStage {
packet_count,
id,
);
inc_new_counter_debug!("banking_stage-transactions_received", packet_count);
let mut proc_start = Measure::start("process_packets_transactions_process");
let mut new_tx_count = 0;
@ -1358,6 +1361,8 @@ impl BankingStage {
let poh_recorder_bank = poh.lock().unwrap().get_poh_recorder_bank();
let working_bank_start = poh_recorder_bank.working_bank_start();
if PohRecorder::get_working_bank_if_not_expired(&working_bank_start).is_none() {
qos_service
.accumulate_tpu_buffered_packets_count(packet_batch.packets.len() as u64);
Self::push_unprocessed(
buffered_packet_batches,
packet_batch,
@ -1378,6 +1383,7 @@ impl BankingStage {
bank_creation_time,
} = &*working_bank_start.unwrap();
qos_service.accumulate_tpu_ingested_packets_count(packet_batch.packets.len() as u64);
let (processed, verified_txs_len, unprocessed_indexes) =
Self::process_packets_transactions(
working_bank,
@ -1392,6 +1398,9 @@ impl BankingStage {
);
new_tx_count += processed;
qos_service.accumulated_verified_txs_count(verified_txs_len as u64);
qos_service.accumulated_processed_txs_count(processed as u64);
qos_service.accumulated_retryable_txs_count(unprocessed_indexes.len() as u64);
// Collect any unprocessed transactions in this batch for forwarding
Self::push_unprocessed(
@ -2272,7 +2281,7 @@ mod tests {
0,
None,
&gossip_vote_sender,
&Arc::new(QosService::new(Arc::new(RwLock::new(CostModel::default())))),
&QosService::new(Arc::new(RwLock::new(CostModel::default())), 1),
)
.0
.unwrap();
@ -2314,7 +2323,7 @@ mod tests {
0,
None,
&gossip_vote_sender,
&Arc::new(QosService::new(Arc::new(RwLock::new(CostModel::default())))),
&QosService::new(Arc::new(RwLock::new(CostModel::default())), 1),
)
.0,
Err(PohRecorderError::MaxHeightReached)
@ -2402,7 +2411,7 @@ mod tests {
0,
None,
&gossip_vote_sender,
&Arc::new(QosService::new(Arc::new(RwLock::new(CostModel::default())))),
&QosService::new(Arc::new(RwLock::new(CostModel::default())), 1),
);
poh_recorder
@ -2511,7 +2520,7 @@ mod tests {
&recorder,
None,
&gossip_vote_sender,
&Arc::new(QosService::new(Arc::new(RwLock::new(CostModel::default())))),
&QosService::new(Arc::new(RwLock::new(CostModel::default())), 1),
);
assert_eq!(processed_transactions_count, 0,);
@ -2606,7 +2615,7 @@ mod tests {
enable_cpi_and_log_storage: false,
}),
&gossip_vote_sender,
&Arc::new(QosService::new(Arc::new(RwLock::new(CostModel::default())))),
&QosService::new(Arc::new(RwLock::new(CostModel::default())), 1),
);
transaction_status_service.join().unwrap();
@ -2737,7 +2746,7 @@ mod tests {
None::<Box<dyn Fn()>>,
&BankingStageStats::default(),
&recorder,
&Arc::new(QosService::new(Arc::new(RwLock::new(CostModel::default())))),
&QosService::new(Arc::new(RwLock::new(CostModel::default())), 1),
);
assert_eq!(
buffered_packet_batches[0].1.len(),
@ -2757,7 +2766,7 @@ mod tests {
None::<Box<dyn Fn()>>,
&BankingStageStats::default(),
&recorder,
&Arc::new(QosService::new(Arc::new(RwLock::new(CostModel::default())))),
&QosService::new(Arc::new(RwLock::new(CostModel::default())), 1),
);
if num_expected_unprocessed == 0 {
assert!(buffered_packet_batches.is_empty())
@ -2823,7 +2832,7 @@ mod tests {
test_fn,
&BankingStageStats::default(),
&recorder,
&Arc::new(QosService::new(Arc::new(RwLock::new(CostModel::default())))),
&QosService::new(Arc::new(RwLock::new(CostModel::default())), 1),
);
// Check everything is correct. All indexes after `interrupted_iteration`

View File

@ -10,12 +10,13 @@ use {
cost_tracker::CostTrackerError,
},
solana_sdk::{
timing::AtomicInterval,
clock::Slot,
transaction::{self, SanitizedTransaction, TransactionError},
},
std::{
sync::{
atomic::{AtomicBool, AtomicU64, Ordering},
mpsc::{channel, Receiver, Sender},
Arc, RwLock,
},
thread::{self, Builder, JoinHandle},
@ -23,13 +24,24 @@ use {
},
};
// QosService is local to each banking thread, each instance of QosService provides services to
// one banking thread.
// It hosts a private thread for async metrics reporting, tagged with banking thredas ID. Banking
// threda calls `report_metrics(&bank)` at end of `process_and_record_tramsaction()`, or any time
// it wants, QosService sends `&bank` to reporting thread via channel, signalling stats to be
// reported if new bank slot has changed.
//
pub struct QosService {
// cost_model instance is owned by validator, shared between replay_stage and
// banking_stage. replay_stage writes the latest on-chain program timings to
// it; banking_stage's qos_service reads that information to calculate
// transaction cost, hence RwLock wrapped.
cost_model: Arc<RwLock<CostModel>>,
// QosService hosts metrics object and a private reporting thread, as well as sender to
// communicate with thread.
report_sender: Sender<Arc<Bank>>,
metrics: Arc<QosServiceMetrics>,
// metrics reporting runs on a private thread
reporting_thread: Option<JoinHandle<()>>,
running_flag: Arc<AtomicBool>,
}
@ -46,16 +58,10 @@ impl Drop for QosService {
}
impl QosService {
pub fn new(cost_model: Arc<RwLock<CostModel>>) -> Self {
Self::new_with_reporting_duration(cost_model, 1000u64)
}
pub fn new_with_reporting_duration(
cost_model: Arc<RwLock<CostModel>>,
reporting_duration_ms: u64,
) -> Self {
pub fn new(cost_model: Arc<RwLock<CostModel>>, id: u32) -> Self {
let (report_sender, report_receiver) = channel();
let running_flag = Arc::new(AtomicBool::new(true));
let metrics = Arc::new(QosServiceMetrics::default());
let metrics = Arc::new(QosServiceMetrics::new(id));
let running_flag_clone = running_flag.clone();
let metrics_clone = metrics.clone();
@ -63,18 +69,21 @@ impl QosService {
Builder::new()
.name("solana-qos-service-metrics-repoting".to_string())
.spawn(move || {
Self::reporting_loop(running_flag_clone, metrics_clone, reporting_duration_ms);
Self::reporting_loop(running_flag_clone, metrics_clone, report_receiver);
})
.unwrap(),
);
Self {
cost_model,
metrics,
reporting_thread,
running_flag,
report_sender,
}
}
// invoke cost_model to calculate cost for the given list of transactions
pub fn compute_transaction_costs<'a>(
&self,
transactions: impl Iterator<Item = &'a SanitizedTransaction>,
@ -147,13 +156,53 @@ impl QosService {
select_results
}
// metrics are reported by bank slot
pub fn report_metrics(&self, bank: Arc<Bank>) {
self.report_sender
.send(bank)
.unwrap_or_else(|err| warn!("qos service report metrics failed: {:?}", err));
}
// metrics accumulating apis
pub fn accumulate_tpu_ingested_packets_count(&self, count: u64) {
self.metrics
.tpu_ingested_packets_count
.fetch_add(count, Ordering::Relaxed);
}
pub fn accumulate_tpu_buffered_packets_count(&self, count: u64) {
self.metrics
.tpu_buffered_packets_count
.fetch_add(count, Ordering::Relaxed);
}
pub fn accumulated_verified_txs_count(&self, count: u64) {
self.metrics
.verified_txs_count
.fetch_add(count, Ordering::Relaxed);
}
pub fn accumulated_processed_txs_count(&self, count: u64) {
self.metrics
.processed_txs_count
.fetch_add(count, Ordering::Relaxed);
}
pub fn accumulated_retryable_txs_count(&self, count: u64) {
self.metrics
.retryable_txs_count
.fetch_add(count, Ordering::Relaxed);
}
fn reporting_loop(
running_flag: Arc<AtomicBool>,
metrics: Arc<QosServiceMetrics>,
reporting_duration_ms: u64,
report_receiver: Receiver<Arc<Bank>>,
) {
while running_flag.load(Ordering::Relaxed) {
metrics.report(reporting_duration_ms);
for bank in report_receiver.try_iter() {
metrics.report(bank.slot());
}
thread::sleep(Duration::from_millis(100));
}
}
@ -161,21 +210,97 @@ impl QosService {
#[derive(Default)]
struct QosServiceMetrics {
last_report: AtomicInterval,
// banking_stage creates one QosService instance per working threads, that is uniquely
// identified by id. This field allows to categorize metrics for gossip votes, TPU votes
// and other transactions.
id: u32,
// aggregate metrics per slot
slot: AtomicU64,
// accumulated number of live packets TPU received from verified receiver for processing.
tpu_ingested_packets_count: AtomicU64,
// accumulated number of live packets TPU put into buffer due to no active bank.
tpu_buffered_packets_count: AtomicU64,
// accumulated number of verified txs, which excludes unsanitized transactions and
// non-vote transactions when in vote-only mode from ingested packets
verified_txs_count: AtomicU64,
// accumulated number of transactions been processed, includes those landed and those to be
// returned (due to AccountInUse, and other QoS related reasons)
processed_txs_count: AtomicU64,
// accumulated number of transactions buffered for retry, often due to AccountInUse and QoS
// reasons, includes retried_txs_per_block_limit_count and retried_txs_per_account_limit_count
retryable_txs_count: AtomicU64,
// accumulated time in micro-sec spent in computing transaction cost. It is the main performance
// overhead introduced by cost_model
compute_cost_time: AtomicU64,
// total nummber of transactions in the reporting period to be computed for theit cost. It is
// usually the number of sanitized transactions leader receives.
compute_cost_count: AtomicU64,
// acumulated time in micro-sec spent in tracking each bank's cost. It is the second part of
// overhead introduced
cost_tracking_time: AtomicU64,
// number of transactions to be included in blocks
selected_txs_count: AtomicU64,
// number of transactions to be queued for retry due to its potential to breach block limit
retried_txs_per_block_limit_count: AtomicU64,
// number of transactions to be queued for retry due to its potential to breach writable
// account limit
retried_txs_per_account_limit_count: AtomicU64,
// number of transactions to be queued for retry due to its account data limits
retried_txs_per_account_data_limit_count: AtomicU64,
}
impl QosServiceMetrics {
pub fn report(&self, report_interval_ms: u64) {
if self.last_report.should_update(report_interval_ms) {
pub fn new(id: u32) -> Self {
QosServiceMetrics {
id,
..QosServiceMetrics::default()
}
}
pub fn report(&self, bank_slot: Slot) {
if bank_slot != self.slot.load(Ordering::Relaxed) {
datapoint_info!(
"qos-service-stats",
("id", self.id as i64, i64),
("bank_slot", bank_slot as i64, i64),
(
"tpu_ingested_packets_count",
self.tpu_ingested_packets_count.swap(0, Ordering::Relaxed) as i64,
i64
),
(
"tpu_buffered_packets_count",
self.tpu_buffered_packets_count.swap(0, Ordering::Relaxed) as i64,
i64
),
(
"verified_txs_count",
self.verified_txs_count.swap(0, Ordering::Relaxed) as i64,
i64
),
(
"processed_txs_count",
self.processed_txs_count.swap(0, Ordering::Relaxed) as i64,
i64
),
(
"retryable_txs_count",
self.retryable_txs_count.swap(0, Ordering::Relaxed) as i64,
i64
),
(
"compute_cost_time",
self.compute_cost_time.swap(0, Ordering::Relaxed) as i64,
@ -215,6 +340,7 @@ impl QosServiceMetrics {
i64
),
);
self.slot.store(bank_slot, Ordering::Relaxed);
}
}
}
@ -259,7 +385,7 @@ mod tests {
let txs = vec![transfer_tx.clone(), vote_tx.clone(), vote_tx, transfer_tx];
let cost_model = Arc::new(RwLock::new(CostModel::default()));
let qos_service = QosService::new(cost_model.clone());
let qos_service = QosService::new(cost_model.clone(), 1);
let txs_costs = qos_service.compute_transaction_costs(txs.iter());
// verify the size of txs_costs and its contents
@ -307,7 +433,7 @@ mod tests {
// make a vec of txs
let txs = vec![transfer_tx.clone(), vote_tx.clone(), transfer_tx, vote_tx];
let qos_service = QosService::new(cost_model);
let qos_service = QosService::new(cost_model, 1);
let txs_costs = qos_service.compute_transaction_costs(txs.iter());
// set cost tracker limit to fit 1 transfer tx, vote tx bypasses limit check
@ -324,83 +450,4 @@ mod tests {
assert!(results[2].is_err());
assert!(results[3].is_ok());
}
#[test]
fn test_async_report_metrics() {
solana_logger::setup();
//solana_logger::setup_with_default("solana=info");
// make a vec of txs
let txs_count = 128usize;
let keypair = Keypair::new();
let transfer_tx = SanitizedTransaction::from_transaction_for_tests(
system_transaction::transfer(&keypair, &keypair.pubkey(), 1, Hash::default()),
);
let mut txs_1 = Vec::with_capacity(txs_count);
let mut txs_2 = Vec::with_capacity(txs_count);
for _i in 0..txs_count {
txs_1.push(transfer_tx.clone());
txs_2.push(transfer_tx.clone());
}
// set reporting duration to long enough so the stats wouldn't reset during testing
let ten_min = 600_000u64;
let cost_model = Arc::new(RwLock::new(CostModel::default()));
let qos_service = Arc::new(QosService::new_with_reporting_duration(cost_model, ten_min));
let qos_service_1 = qos_service.clone();
let qos_service_2 = qos_service.clone();
let th_1 = Builder::new()
.name("test-producer-1".to_string())
.spawn(move || {
debug!("thread 1 starts with {} txs", txs_1.len());
let tx_costs = qos_service_1.compute_transaction_costs(txs_1.iter());
assert_eq!(txs_count, tx_costs.len());
debug!(
"thread 1 done, generated {} count, see service count as {}",
txs_count,
qos_service_1
.metrics
.compute_cost_count
.load(Ordering::Relaxed)
);
})
.unwrap();
let th_2 = Builder::new()
.name("test-producer-2".to_string())
.spawn(move || {
debug!("thread 2 starts with {} txs", txs_2.len());
let tx_costs = qos_service_2.compute_transaction_costs(txs_2.iter());
assert_eq!(txs_count, tx_costs.len());
debug!(
"thread 2 done, generated {} count, see service count as {}",
txs_count,
qos_service_2
.metrics
.compute_cost_count
.load(Ordering::Relaxed)
);
})
.unwrap();
th_1.join().expect("qos service 1 panicked");
th_2.join().expect("qos service 2 panicked");
debug!(
"all threads joined. count {}",
qos_service
.metrics
.compute_cost_count
.load(Ordering::Relaxed)
);
assert_eq!(
txs_count as u64 * 2,
qos_service
.metrics
.compute_cost_count
.load(Ordering::Relaxed)
);
}
}