Faster dedup v1.8 (solana-labs#22619)

* Faster dedup

Author: aeyakovenko

Cargo.lock

@@ -7,13 +7,11 @@
 
 use {
     crate::sigverify,
-    core::time::Duration,
     crossbeam_channel::{SendError, Sender as CrossbeamSender},
     itertools::Itertools,
-    solana_bloom::bloom::{AtomicBloom, Bloom},
     solana_measure::measure::Measure,
     solana_perf::packet::PacketBatch,
-    solana_perf::sigverify::dedup_packets,
+    solana_perf::sigverify::Deduper,
     solana_sdk::timing,
     solana_streamer::streamer::{self, PacketBatchReceiver, StreamerError},
     std::{
@@ -215,7 +213,7 @@ impl SigVerifyStage {
     }
 
     fn verifier<T: SigVerifier>(
-        bloom: &AtomicBloom<&[u8]>,
+        deduper: &Deduper,
         recvr: &PacketBatchReceiver,
         sendr: &CrossbeamSender<Vec<PacketBatch>>,
         verifier: &T,
@@ -231,15 +229,15 @@ impl SigVerifyStage {
         );
 
         let mut dedup_time = Measure::start("sigverify_dedup_time");
-        let dedup_fail = dedup_packets(bloom, &mut batches) as usize;
+        let dedup_fail = deduper.dedup_packets(&mut batches) as usize;
         dedup_time.stop();
-        let valid_packets = num_packets.saturating_sub(dedup_fail);
+        let num_unique = num_packets.saturating_sub(dedup_fail);
 
         let mut discard_time = Measure::start("sigverify_discard_time");
-        if valid_packets > MAX_SIGVERIFY_BATCH {
+        if num_unique > MAX_SIGVERIFY_BATCH {
             Self::discard_excess_packets(&mut batches, MAX_SIGVERIFY_BATCH)
         };
-        let excess_fail = valid_packets.saturating_sub(MAX_SIGVERIFY_BATCH);
+        let excess_fail = num_unique.saturating_sub(MAX_SIGVERIFY_BATCH);
         discard_time.stop();
 
         let mut verify_batch_time = Measure::start("sigverify_batch_time");
@@ -298,25 +296,16 @@ impl SigVerifyStage {
         let verifier = verifier.clone();
         let mut stats = SigVerifierStats::default();
         let mut last_print = Instant::now();
-        const MAX_BLOOM_AGE: Duration = Duration::from_millis(2_000);
-        const MAX_BLOOM_ITEMS: usize = 1_000_000;
-        const MAX_BLOOM_FAIL: f64 = 0.0001;
-        const MAX_BLOOM_BITS: usize = 8 << 22;
+        const MAX_DEDUPER_AGE_MS: u64 = 2_000;
+        const MAX_DEDUPER_ITEMS: u32 = 1_000_000;
         Builder::new()
             .name("solana-verifier".to_string())
             .spawn(move || {
-                let mut bloom =
-                    Bloom::random(MAX_BLOOM_ITEMS, MAX_BLOOM_FAIL, MAX_BLOOM_BITS).into();
-                let mut bloom_age = Instant::now();
+                let mut deduper = Deduper::new(MAX_DEDUPER_ITEMS, MAX_DEDUPER_AGE_MS);
                 loop {
-                    let now = Instant::now();
-                    if now.duration_since(bloom_age) > MAX_BLOOM_AGE {
-                        bloom =
-                            Bloom::random(MAX_BLOOM_ITEMS, MAX_BLOOM_FAIL, MAX_BLOOM_BITS).into();
-                        bloom_age = now;
-                    }
+                    deduper.reset();
                     if let Err(e) = Self::verifier(
-                        &bloom,
+                        &deduper,
                         &packet_receiver,
                         &verified_sender,
                         &verifier,

core/src/sigverify_stage.rs

@@ -12,6 +12,7 @@ edition = "2018"
 [dependencies]
 bincode = "1.3.1"
 curve25519-dalek = { version = "2" }
+ahash = "0.7.6"
 dlopen = "0.1.8"
 dlopen_derive = "0.1.4"
 lazy_static = "1.4.0"

perf/Cargo.toml

@@ -5,14 +5,15 @@ extern crate test;
 
 use {
     rand::prelude::*,
-    solana_bloom::bloom::{AtomicBloom, Bloom},
     solana_perf::{
         packet::{to_packet_batches, PacketBatch},
         sigverify,
     },
     test::Bencher,
 };
 
+const NUM: usize = 4096;
+
 fn test_packet_with_size(size: usize, rng: &mut ThreadRng) -> Vec<u8> {
     // subtract 8 bytes because the length will get serialized as well
     (0..size.checked_sub(8).unwrap())
@@ -22,19 +23,14 @@ fn test_packet_with_size(size: usize, rng: &mut ThreadRng) -> Vec<u8> {
 
 fn do_bench_dedup_packets(bencher: &mut Bencher, mut batches: Vec<PacketBatch>) {
     // verify packets
-    let mut bloom: AtomicBloom<&[u8]> = Bloom::random(1_000_000, 0.0001, 8 << 22).into();
+    let mut deduper = sigverify::Deduper::new(1_000_000, 2_000);
     bencher.iter(|| {
-        // bench
-        sigverify::dedup_packets(&bloom, &mut batches);
-
-        // reset
-        bloom.clear_for_tests();
-        batches.iter_mut().for_each(|batch| {
-            batch.packets.iter_mut().for_each(|p| {
-                p.meta.discard = false;
-            })
-        });
-    })
+        let _ans = deduper.dedup_packets(&mut batches);
+        deduper.reset();
+        batches
+            .iter_mut()
+            .for_each(|b| b.packets.iter_mut().for_each(|p| p.meta.discard = false));
+    });
 }
 
 #[bench]
@@ -45,7 +41,7 @@ fn bench_dedup_same_small_packets(bencher: &mut Bencher) {
 
     let batches = to_packet_batches(
         &std::iter::repeat(small_packet)
-            .take(4096)
+            .take(NUM)
             .collect::<Vec<_>>(),
         128,
     );
@@ -60,7 +56,7 @@ fn bench_dedup_same_big_packets(bencher: &mut Bencher) {
     let big_packet = test_packet_with_size(1024, &mut rng);
 
     let batches = to_packet_batches(
-        &std::iter::repeat(big_packet).take(4096).collect::<Vec<_>>(),
+        &std::iter::repeat(big_packet).take(NUM).collect::<Vec<_>>(),
         128,
     );
 
@@ -73,7 +69,7 @@ fn bench_dedup_diff_small_packets(bencher: &mut Bencher) {
     let mut rng = rand::thread_rng();
 
     let batches = to_packet_batches(
-        &(0..4096)
+        &(0..NUM)
             .map(|_| test_packet_with_size(128, &mut rng))
             .collect::<Vec<_>>(),
         128,
@@ -88,11 +84,35 @@ fn bench_dedup_diff_big_packets(bencher: &mut Bencher) {
     let mut rng = rand::thread_rng();
 
     let batches = to_packet_batches(
-        &(0..4096)
+        &(0..NUM)
             .map(|_| test_packet_with_size(1024, &mut rng))
             .collect::<Vec<_>>(),
         128,
     );
 
     do_bench_dedup_packets(bencher, batches);
 }
+
+#[bench]
+#[ignore]
+fn bench_dedup_baseline(bencher: &mut Bencher) {
+    let mut rng = rand::thread_rng();
+
+    let batches = to_packet_batches(
+        &(0..0)
+            .map(|_| test_packet_with_size(128, &mut rng))
+            .collect::<Vec<_>>(),
+        128,
+    );
+
+    do_bench_dedup_packets(bencher, batches);
+}
+
+#[bench]
+#[ignore]
+fn bench_dedup_reset(bencher: &mut Bencher) {
+    let mut deduper = sigverify::Deduper::new(1_000_000, 0);
+    bencher.iter(|| {
+        deduper.reset();
+    });
+}