Dedup bloom filter is too slow (#22607)

* Faster dedup

* use ahash

* fixup

* single threaded

* use duration type

* remove the count

* fixup

perf/benches/dedup.rs

@@ -5,14 +5,16 @@ extern crate test;
 
 use {
     rand::prelude::*,
-    solana_bloom::bloom::{AtomicBloom, Bloom},
     solana_perf::{
         packet::{to_packet_batches, PacketBatch},
         sigverify,
     },
+    std::time::Duration,
     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())
@@ -22,20 +24,14 @@ 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, Duration::from_millis(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.set_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.set_discard(false)));
+    });
 }
 
 #[bench]
@@ -46,7 +42,7 @@ 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,
     );
@@ -61,7 +57,7 @@ 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,
     );
 
@@ -74,7 +70,7 @@ 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,
@@ -89,7 +85,7 @@ 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,
@@ -97,3 +93,27 @@ fn bench_dedup_diff_big_packets(bencher: &mut Bencher) {
 
     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, Duration::from_millis(0));
+    bencher.iter(|| {
+        deduper.reset();
+    });
+}