verification-rate based thread scaling

ethcore/src/verification/queue/mod.rs

@@ -18,7 +18,7 @@
 //! Sorts them ready for blockchain insertion.
 
 use std::thread::{JoinHandle, self};
-use std::sync::atomic::{AtomicBool, Ordering as AtomicOrdering};
+use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering as AtomicOrdering};
 use std::sync::{Condvar as SCondvar, Mutex as SMutex};
 use util::*;
 use io::*;
@@ -113,7 +113,7 @@ pub struct VerificationQueue<K: Kind> {
 	ready_signal: Arc<QueueSignal>,
 	empty: Arc<SCondvar>,
 	processing: RwLock<HashSet<H256>>,
-	rolling_sample: Mutex<VecDeque<(usize, usize)>>,
+	ticks_since_adjustment: AtomicUsize,
 	max_queue_size: usize,
 	max_mem_use: usize,
 }
@@ -152,6 +152,8 @@ struct Verification<K: Kind> {
 	bad: Mutex<HashSet<H256>>,
 	more_to_verify: SMutex<()>,
 	empty: SMutex<()>,
+	verified_count: AtomicUsize,
+	drained: AtomicUsize,
 }
 
 impl<K: Kind> VerificationQueue<K> {
@@ -164,7 +166,8 @@ impl<K: Kind> VerificationQueue<K> {
 			bad: Mutex::new(HashSet::new()),
 			more_to_verify: SMutex::new(()),
 			empty: SMutex::new(()),
-
+			verified_count: AtomicUsize::new(0),
+			drained: AtomicUsize::new(0),
 		});
 		let more_to_verify = Arc::new(SCondvar::new());
 		let deleting = Arc::new(AtomicBool::new(false));
@@ -186,7 +189,7 @@ impl<K: Kind> VerificationQueue<K> {
 			deleting: deleting,
 			processing: RwLock::new(HashSet::new()),
 			empty: empty,
-			rolling_sample: Mutex::new(VecDeque::new()),
+			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),
 		};
@@ -248,7 +251,7 @@ impl<K: Kind> VerificationQueue<K> {
 						// we're next!
 						let mut verified = verification.verified.lock();
 						let mut bad = verification.bad.lock();
-						VerificationQueue::drain_verifying(&mut verifying, &mut verified, &mut bad);
+						VerificationQueue::drain_verifying(&mut verifying, &mut verified, &mut bad, &verification.verified_count);
 						ready.set();
 					}
 				},
@@ -261,7 +264,7 @@ impl<K: Kind> VerificationQueue<K> {
 					verifying.retain(|e| e.hash != hash);
 
 					if verifying.front().map_or(false, |x| x.output.is_some()) {
-						VerificationQueue::drain_verifying(&mut verifying, &mut verified, &mut bad);
+						VerificationQueue::drain_verifying(&mut verifying, &mut verified, &mut bad, &verification.verified_count);
 						ready.set();
 					}
 				}
@@ -269,7 +272,13 @@ impl<K: Kind> VerificationQueue<K> {
 		}
 	}
 
-	fn drain_verifying(verifying: &mut VecDeque<Verifying<K>>, verified: &mut VecDeque<K::Verified>, bad: &mut HashSet<H256>) {
+	fn drain_verifying(
+		verifying: &mut VecDeque<Verifying<K>>,
+		verified: &mut VecDeque<K::Verified>,
+		bad: &mut HashSet<H256>,
+		v_count: &AtomicUsize
+	) {
+		let start_len = verified.len();
 		while let Some(output) = verifying.front_mut().and_then(|x| x.output.take()) {
 			assert!(verifying.pop_front().is_some());
 
@@ -279,6 +288,8 @@ impl<K: Kind> VerificationQueue<K> {
 				verified.push_back(output);
 			}
 		}
+
+		v_count.fetch_add(verified.len() - start_len, AtomicOrdering::AcqRel);
 	}
 
 	/// Clear the queue and stop verification activity.
@@ -389,6 +400,8 @@ impl<K: Kind> VerificationQueue<K> {
 		let count = min(max, verified.len());
 		let result = verified.drain(..count).collect::<Vec<_>>();
 
+		self.verification.drained.fetch_add(count, AtomicOrdering::AcqRel);
+
 		self.ready_signal.reset();
 		if !verified.is_empty() {
 			self.ready_signal.set();
@@ -429,53 +442,49 @@ impl<K: Kind> VerificationQueue<K> {
 	/// Optimise memory footprint of the heap fields, and adjust the number of threads
 	/// to better suit the workload.
 	pub fn collect_garbage(&self) {
-		// thresholds for adding and removing verifier threads
-		// these are unbalanced since having all blocks verified
-		// is the desirable position.
-		const ADD_THREAD_THRESHOLD: usize = 10;
-		const DEL_THREAD_THRESHOLD: usize = 20;
-
 		// number of ticks to average queue stats over
 		// when deciding whether to change the number of verifiers.
-		const SAMPLE_SIZE: usize = 5;
+		const READJUSTMENT_PERIOD: usize = 5;
-
-		let (u_len, v_len) = {
-			let u_len = {
-				let mut v = self.verification.unverified.lock();
-				v.shrink_to_fit();
-				v.len()
-			};
 
+		{
+			self.verification.unverified.lock().shrink_to_fit();
 			self.verification.verifying.lock().shrink_to_fit();
+			self.verification.verified.lock().shrink_to_fit();
+		}
 
-			let v_len = {
-				let mut v = self.verification.verified.lock();
-				v.shrink_to_fit();
-				v.len()
-			};
-
-			(u_len, v_len)
-		};
 		self.processing.write().shrink_to_fit();
 
-		let (u_len, v_len) = {
+		if self.ticks_since_adjustment.load(AtomicOrdering::SeqCst) == READJUSTMENT_PERIOD {
-			let mut sample = self.rolling_sample.lock();
+			self.ticks_since_adjustment.store(0, AtomicOrdering::SeqCst);
-			sample.push_back((u_len, v_len));
+		} else {
+			self.ticks_since_adjustment.fetch_add(1, AtomicOrdering::SeqCst);
+			return;
+		}
 
-			if sample.len() > SAMPLE_SIZE {
+		let v_count = self.verification.verified_count.load(AtomicOrdering::Acquire);
-				let _ = sample.pop_front();
+		let drained = self.verification.drained.load(AtomicOrdering::Acquire);
-			}
 
-			sample.iter().cloned().fold((0, 0), |(u_t, v_t), (u_i, v_i)| (u_t + u_i, v_t + v_i))
+		self.verification.verified_count.store(0, AtomicOrdering::Release);
+		self.verification.drained.store(0, AtomicOrdering::Release);
+
+		// compute the average rate of verification per thread and determine
+		// how many are necessary to match the rate of draining.
+		let num_verifiers = self.verifiers.lock().len();
+		let v_count_per = v_count as f64 / num_verifiers as f64;
+		let needed = if v_count < 20 {
+			1
+		} else {
+			(drained as f64 / v_count_per as f64).ceil() as usize
 		};
 
-		// more than 10x as many unverified as verified.
+		trace!(target: "verification", "v_rate_per={}, drained={}, scaling to {} verifiers",
-		if v_len * ADD_THREAD_THRESHOLD < u_len {
+			v_count_per, drained, needed);
+
+		for _ in num_verifiers..needed {
 			self.add_verifier();
 		}
 
-		// more than 20x as many verified as unverified.
+		for _ in needed..num_verifiers {
-		if u_len * DEL_THREAD_THRESHOLD < v_len {
 			self.remove_verifier();
 		}
 	}