Merge pull request #2445 from ethcore/adaptive_queue_threads

Use an adaptive number of threads in the verification queue

ethcore/src/verification/queue/mod.rs

@@ -35,6 +35,9 @@ pub mod kind;
 const MIN_MEM_LIMIT: usize = 16384;
 const MIN_QUEUE_LIMIT: usize = 512;
 
+// maximum possible number of verification threads.
+const MAX_VERIFIERS: usize = 8;
+
 /// Type alias for block queue convenience.
 pub type BlockQueue = VerificationQueue<self::kind::Blocks>;
 
@@ -61,6 +64,37 @@ impl Default for Config {
 	}
 }
 
+struct VerifierHandle {
+	deleting: Arc<AtomicBool>,
+	sleep: Arc<AtomicBool>,
+	thread: JoinHandle<()>,
+}
+
+impl VerifierHandle {
+	// signal to the verifier thread that it should sleep.
+	fn sleep(&self) {
+		self.sleep.store(true, AtomicOrdering::SeqCst);
+	}
+
+	// signal to the verifier thread that it should wake up.
+	fn wake_up(&self) {
+		self.sleep.store(false, AtomicOrdering::SeqCst);
+		self.thread.thread().unpark();
+	}
+
+	// signal to the verifier thread that it should conclude its
+	// operations.
+	fn conclude(&self) {
+		self.wake_up();
+		self.deleting.store(true, AtomicOrdering::Release);
+	}
+
+	// join the verifier thread.
+	fn join(self) {
+		self.thread.join().expect("Verifier thread panicked");
+	}
+}
+
 /// An item which is in the process of being verified.
 pub struct Verifying<K: Kind> {
 	hash: H256,
@@ -97,11 +131,12 @@ pub struct VerificationQueue<K: Kind> {
 	engine: Arc<Engine>,
 	more_to_verify: Arc<SCondvar>,
 	verification: Arc<Verification<K>>,
-	verifiers: Vec<JoinHandle<()>>,
+	verifiers: Mutex<(Vec<VerifierHandle>, usize)>,
 	deleting: Arc<AtomicBool>,
 	ready_signal: Arc<QueueSignal>,
 	empty: Arc<SCondvar>,
 	processing: RwLock<HashSet<H256>>,
+	ticks_since_adjustment: AtomicUsize,
 	max_queue_size: usize,
 	max_mem_use: usize,
 }
@@ -157,6 +192,7 @@ struct Verification<K: Kind> {
 	more_to_verify: SMutex<()>,
 	empty: SMutex<()>,
 	sizes: Sizes,
+	check_seal: bool,
 }
 
 impl<K: Kind> VerificationQueue<K> {
@@ -173,7 +209,8 @@ impl<K: Kind> VerificationQueue<K> {
 				unverified: AtomicUsize::new(0),
 				verifying: AtomicUsize::new(0),
 				verified: AtomicUsize::new(0),
-			}
+			},
+			check_seal: check_seal,
 		});
 		let more_to_verify = Arc::new(SCondvar::new());
 		let deleting = Arc::new(AtomicBool::new(false));
@@ -185,44 +222,82 @@ impl<K: Kind> VerificationQueue<K> {
 		let empty = Arc::new(SCondvar::new());
 		let panic_handler = PanicHandler::new_in_arc();
 
-		let mut verifiers: Vec<JoinHandle<()>> = Vec::new();
-		let thread_count = max(::num_cpus::get(), 3) - 2;
-		for i in 0..thread_count {
-			let verification = verification.clone();
-			let engine = engine.clone();
-			let more_to_verify = more_to_verify.clone();
-			let ready_signal = ready_signal.clone();
-			let empty = empty.clone();
+		let max_verifiers = min(::num_cpus::get(), MAX_VERIFIERS);
+		let default_amount = max(::num_cpus::get(), 3) - 2;
+		let mut verifiers = Vec::with_capacity(max_verifiers);
+
+		debug!(target: "verification", "Allocating {} verifiers, {} initially active", max_verifiers, default_amount);
+
+		for i in 0..max_verifiers {
+			debug!(target: "verification", "Adding verification thread #{}", i);
+
 			let deleting = deleting.clone();
 			let panic_handler = panic_handler.clone();
-			verifiers.push(
-				thread::Builder::new()
-				.name(format!("Verifier #{}", i))
-				.spawn(move || {
-					panic_handler.catch_panic(move || {
-						VerificationQueue::verify(verification, engine, more_to_verify, ready_signal, deleting, empty, check_seal)
-					}).unwrap()
-				})
-				.expect("Error starting block verification thread")
-			);
+			let verification = verification.clone();
+			let engine = engine.clone();
+			let wait = more_to_verify.clone();
+			let ready = ready_signal.clone();
+			let empty = empty.clone();
+
+			// enable only the first few verifiers.
+			let sleep = if i < default_amount {
+				Arc::new(AtomicBool::new(false))
+			} else {
+				Arc::new(AtomicBool::new(true))
+			};
+
+			verifiers.push(VerifierHandle {
+				deleting: deleting.clone(),
+				sleep: sleep.clone(),
+				thread: thread::Builder::new()
+					.name(format!("Verifier #{}", i))
+					.spawn(move || {
+						panic_handler.catch_panic(move || {
+							VerificationQueue::verify(verification, engine, wait, ready, deleting, empty, sleep)
+						}).unwrap()
+					})
+					.expect("Failed to create verifier thread.")
+			});
 		}
+
 		VerificationQueue {
 			engine: engine,
 			panic_handler: panic_handler,
-			ready_signal: ready_signal.clone(),
-			more_to_verify: more_to_verify.clone(),
-			verification: verification.clone(),
-			verifiers: verifiers,
-			deleting: deleting.clone(),
+			ready_signal: ready_signal,
+			more_to_verify: more_to_verify,
+			verification: verification,
+			verifiers: Mutex::new((verifiers, default_amount)),
+			deleting: deleting,
 			processing: RwLock::new(HashSet::new()),
-			empty: empty.clone(),
+			empty: empty,
+			ticks_since_adjustment: AtomicUsize::new(0),
 			max_queue_size: max(config.max_queue_size, MIN_QUEUE_LIMIT),
 			max_mem_use: max(config.max_mem_use, MIN_MEM_LIMIT),
 		}
 	}
 
-	fn verify(verification: Arc<Verification<K>>, engine: Arc<Engine>, wait: Arc<SCondvar>, ready: Arc<QueueSignal>, deleting: Arc<AtomicBool>, empty: Arc<SCondvar>, check_seal: bool) {
+	fn verify(
+		verification: Arc<Verification<K>>,
+		engine: Arc<Engine>,
+		wait: Arc<SCondvar>,
+		ready: Arc<QueueSignal>,
+		deleting: Arc<AtomicBool>,
+		empty: Arc<SCondvar>,
+		sleep: Arc<AtomicBool>,
+	) {
 		while !deleting.load(AtomicOrdering::Acquire) {
+			{
+				while sleep.load(AtomicOrdering::SeqCst) {
+					trace!(target: "verification", "Verifier sleeping");
+					::std::thread::park();
+					trace!(target: "verification", "Verifier waking up");
+
+					if deleting.load(AtomicOrdering::Acquire) {
+						return;
+					}
+				}
+			}
+
 			{
 				let mut more_to_verify = verification.more_to_verify.lock().unwrap();
 
@@ -255,7 +330,7 @@ impl<K: Kind> VerificationQueue<K> {
 			};
 
 			let hash = item.hash();
-			let is_ready = match K::verify(item, &*engine, check_seal) {
+			let is_ready = match K::verify(item, &*engine, verification.check_seal) {
 				Ok(verified) => {
 					let mut verifying = verification.verifying.lock();
 					let mut idx = None;
@@ -302,9 +377,15 @@ impl<K: Kind> VerificationQueue<K> {
 		}
 	}
 
-	fn drain_verifying(verifying: &mut VecDeque<Verifying<K>>, verified: &mut VecDeque<K::Verified>, bad: &mut HashSet<H256>, sizes: &Sizes) {
+	fn drain_verifying(
+		verifying: &mut VecDeque<Verifying<K>>,
+		verified: &mut VecDeque<K::Verified>,
+		bad: &mut HashSet<H256>,
+		sizes: &Sizes,
+	) {
 		let mut removed_size = 0;
 		let mut inserted_size = 0;
+
 		while let Some(output) = verifying.front_mut().and_then(|x| x.output.take()) {
 			assert!(verifying.pop_front().is_some());
 			let size = output.heap_size_of_children();
@@ -487,14 +568,85 @@ impl<K: Kind> VerificationQueue<K> {
 		}
 	}
 
-	/// Optimise memory footprint of the heap fields.
+	/// Optimise memory footprint of the heap fields, and adjust the number of threads
+	/// to better suit the workload.
 	pub fn collect_garbage(&self) {
-		{
-			self.verification.unverified.lock().shrink_to_fit();
+		// number of ticks to average queue stats over
+		// when deciding whether to change the number of verifiers.
+		#[cfg(not(test))]
+		const READJUSTMENT_PERIOD: usize = 12;
+
+		#[cfg(test)]
+		const READJUSTMENT_PERIOD: usize = 1;
+
+		let (u_len, v_len) = {
+			let u_len = {
+				let mut q = self.verification.unverified.lock();
+				q.shrink_to_fit();
+				q.len()
+			};
 			self.verification.verifying.lock().shrink_to_fit();
-			self.verification.verified.lock().shrink_to_fit();
-		}
+
+			let v_len = {
+				let mut q = self.verification.verified.lock();
+				q.shrink_to_fit();
+				q.len()
+			};
+
+			(u_len as isize, v_len as isize)
+		};
+
 		self.processing.write().shrink_to_fit();
+
+		if self.ticks_since_adjustment.fetch_add(1, AtomicOrdering::SeqCst) + 1 >= READJUSTMENT_PERIOD {
+			self.ticks_since_adjustment.store(0, AtomicOrdering::SeqCst);
+		} else {
+			return;
+		}
+
+		let current = self.verifiers.lock().1;
+
+		let diff = (v_len - u_len).abs();
+		let total = v_len + u_len;
+
+		self.scale_verifiers(
+			if u_len < 20 {
+				1
+			} else if diff <= total / 10 {
+				current
+			} else if v_len > u_len {
+				current - 1
+			} else {
+				current + 1
+			}
+		);
+	}
+
+	// wake up or sleep verifiers to get as close to the target as
+	// possible, never going over the amount of initially allocated threads
+	// or below 1.
+	fn scale_verifiers(&self, target: usize) {
+		let mut verifiers = self.verifiers.lock();
+		let &mut (ref mut verifiers, ref mut verifier_count) = &mut *verifiers;
+
+		let target = min(verifiers.len(), target);
+		let target = max(1, target);
+
+		debug!(target: "verification", "Scaling from {} to {} verifiers", verifier_count, target);
+
+		// scaling up
+		for i in *verifier_count..target {
+			debug!(target: "verification", "Waking up verifier {}", i);
+			verifiers[i].wake_up();
+		}
+
+		// scaling down.
+		for i in target..*verifier_count {
+			debug!(target: "verification", "Putting verifier {} to sleep", i);
+			verifiers[i].sleep();
+		}
+
+		*verifier_count = target;
 	}
 }
 
@@ -509,10 +661,23 @@ impl<K: Kind> Drop for VerificationQueue<K> {
 		trace!(target: "shutdown", "[VerificationQueue] Closing...");
 		self.clear();
 		self.deleting.store(true, AtomicOrdering::Release);
-		self.more_to_verify.notify_all();
-		for t in self.verifiers.drain(..) {
-			t.join().unwrap();
+
+		let mut verifiers = self.verifiers.get_mut();
+		let mut verifiers = &mut verifiers.0;
+
+		// first pass to signal conclusion. must be done before
+		// notify or deadlock possible.
+		for handle in verifiers.iter() {
+			handle.conclude();
 		}
+
+		self.more_to_verify.notify_all();
+
+		// second pass to join.
+		for handle in verifiers.drain(..) {
+			handle.join();
+		}
+
 		trace!(target: "shutdown", "[VerificationQueue] Closed.");
 	}
 }
@@ -611,4 +776,56 @@ mod tests {
 		}
 		assert!(queue.queue_info().is_full());
 	}
+
+	#[test]
+	fn scaling_limits() {
+		use super::MAX_VERIFIERS;
+
+		let queue = get_test_queue();
+		queue.scale_verifiers(MAX_VERIFIERS + 1);
+
+		assert!(queue.verifiers.lock().1 < MAX_VERIFIERS + 1);
+
+		queue.scale_verifiers(0);
+
+		assert!(queue.verifiers.lock().1 == 1);
+	}
+
+	#[test]
+	fn readjust_verifiers() {
+		let queue = get_test_queue();
+
+		// put all the verifiers to sleep to ensure 
+		// the test isn't timing sensitive.
+		let num_verifiers = {
+			let verifiers = queue.verifiers.lock();
+			for i in 0..verifiers.1 {
+				verifiers.0[i].sleep();
+			}
+
+			verifiers.1
+		};
+
+		for block in get_good_dummy_block_seq(5000) {
+			queue.import(Unverified::new(block)).expect("Block good by definition; qed");
+		}
+
+		// almost all unverified == bump verifier count.
+		queue.collect_garbage();
+		assert_eq!(queue.verifiers.lock().1, num_verifiers + 1);
+
+		// wake them up again and verify everything.
+		{
+			let verifiers = queue.verifiers.lock();
+			for i in 0..verifiers.1 {
+				verifiers.0[i].wake_up();
+			}
+		}
+
+		queue.flush();
+
+		// nothing to verify == use minimum number of verifiers.
+		queue.collect_garbage();
+		assert_eq!(queue.verifiers.lock().1, 1);
+	}
 }