perf(memory): use thread-local sequence-based memory eviction policy (#…

…16087)

Signed-off-by: MrCroxx <[email protected]>

grafana/risingwave-dev-dashboard.dashboard.py

@@ -3697,22 +3697,26 @@ def section_memory_manager(outer_panels):
                         ),
                     ],
                 ),
-                panels.timeseries_count(
-                    "LRU manager watermark steps",
+                panels.timeseries(
+                    "LRU manager eviction policy",
                     "",
                     [
                         panels.target(
-                            f"{metric('lru_watermark_step')}",
+                            f"{metric('lru_eviction_policy')}",
                             "",
                         ),
                     ],
                 ),
-                panels.timeseries_ms(
-                    "LRU manager diff between watermark_time and now (ms)",
-                    "watermark_time is the current lower watermark of cached data. physical_now is the current time of the machine. The diff (physical_now - watermark_time) shows how much data is cached.",
+                panels.timeseries(
+                    "LRU manager sequence",
+                    "",
                     [
                         panels.target(
-                            f"{metric('lru_physical_now_ms')} - {metric('lru_current_watermark_time_ms')}",
+                            f"{metric('lru_latest_sequence')}",
+                            "",
+                        ),
+                        panels.target(
+                            f"{metric('lru_watermark_sequence')}",
                             "",
                         ),
                     ],

src/common/Cargo.toml

@@ -14,6 +14,7 @@ ignored = ["workspace-hack"]
 normal = ["workspace-hack"]
 
 [dependencies]
+ahash = "0.8"
 anyhow = "1"
 arc-swap = "1"
 arrow-array = { workspace = true }
@@ -48,6 +49,7 @@ fixedbitset = "0.5"
 foyer = { workspace = true }
 futures = { version = "0.3", default-features = false, features = ["alloc"] }
 governor = { version = "0.6", default-features = false, features = ["std"] }
+hashbrown = "0.14"
 hex = "0.4.3"
 http = "0.2"
 humantime = "2.1"
@@ -135,6 +137,7 @@ libc = "0.2"
 mach2 = "0.4"
 
 [dev-dependencies]
+coarsetime = "0.1"
 criterion = { workspace = true }
 expect-test = "1"
 more-asserts = "0.3"
@@ -167,5 +170,13 @@ harness = false
 name = "bench_array"
 harness = false
 
+[[bench]]
+name = "bench_sequencer"
+harness = false
+
+[[bench]]
+name = "bench_lru"
+harness = false
+
 [lints]
 workspace = true

src/common/benches/bench_lru.rs

@@ -0,0 +1,88 @@
+// Copyright 2024 RisingWave Labs
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::hint::black_box;
+use std::sync::atomic::Ordering;
+use std::time::{Duration, Instant};
+
+use itertools::Itertools;
+use lru::LruCache;
+use risingwave_common::lru::LruCache as RwLruCache;
+use risingwave_common::sequence::SEQUENCE_GLOBAL;
+
+fn lru(loops: usize, evict_ratio: u64) -> (usize, Duration) {
+    let mut lru = LruCache::unbounded();
+    let mut evicted = 0;
+    let now = Instant::now();
+    for i in 0..loops as u64 {
+        if i % evict_ratio == 0 && i != 0 {
+            lru.update_epoch(i);
+            while lru.pop_lru_by_epoch(i).is_some() {
+                evicted += 1;
+            }
+        }
+        lru.put(i, i);
+    }
+
+    (evicted, now.elapsed())
+}
+
+fn rw_lru(loops: usize, evict_ratio: u64) -> (usize, Duration) {
+    let mut lru = RwLruCache::unbounded();
+    let mut evicted = 0;
+    let now = Instant::now();
+    for i in 0..loops as u64 {
+        if i % evict_ratio == 0 {
+            let sequence = SEQUENCE_GLOBAL.load(Ordering::Relaxed);
+            while lru.pop_with_sequence(sequence).is_some() {
+                evicted += 1;
+            }
+        }
+        lru.put(i, i);
+    }
+
+    (evicted, now.elapsed())
+}
+
+fn benchmark<F>(name: &str, threads: usize, loops: usize, f: F)
+where
+    F: Fn() -> (usize, Duration) + Clone + Send + 'static,
+{
+    let handles = (0..threads)
+        .map(|_| std::thread::spawn(black_box(f.clone())))
+        .collect_vec();
+    let mut dur = Duration::from_nanos(0);
+    let mut evicted = 0;
+    for handle in handles {
+        let (e, d) = handle.join().unwrap();
+        evicted += e;
+        dur += d;
+    }
+    println!(
+        "{:20} {} threads {} loops: {:?} per iter, total evicted: {}",
+        name,
+        threads,
+        loops,
+        Duration::from_nanos((dur.as_nanos() / threads as u128 / loops as u128) as u64),
+        evicted,
+    );
+}
+
+fn main() {
+    for threads in [1, 4, 8, 16, 32, 64] {
+        println!();
+        benchmark("lru - 1024", threads, 1000000, || lru(1000000, 1024));
+        benchmark("rw  - 1024", threads, 1000000, || rw_lru(1000000, 1024));
+    }
+}

src/common/benches/bench_sequencer.rs

@@ -0,0 +1,170 @@
+// Copyright 2024 RisingWave Labs
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#![feature(lint_reasons)]
+
+use std::cell::RefCell;
+use std::hint::black_box;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+
+use itertools::Itertools;
+use risingwave_common::sequence::*;
+
+thread_local! {
+    pub static SEQUENCER_64_8: RefCell<Sequencer> = const { RefCell::new(Sequencer::new(64, 64 * 8)) };
+    pub static SEQUENCER_64_16: RefCell<Sequencer> = const { RefCell::new(Sequencer::new(64, 64 * 16)) };
+    pub static SEQUENCER_64_32: RefCell<Sequencer> = const { RefCell::new(Sequencer::new(64, 64 * 32)) };
+    pub static SEQUENCER_128_8: RefCell<Sequencer> = const { RefCell::new(Sequencer::new(128, 128 * 8)) };
+    pub static SEQUENCER_128_16: RefCell<Sequencer> = const { RefCell::new(Sequencer::new(128, 128 * 16)) };
+    pub static SEQUENCER_128_32: RefCell<Sequencer> = const { RefCell::new(Sequencer::new(128, 128 * 32)) };
+}
+
+fn coarse(loops: usize) -> Duration {
+    let now = Instant::now();
+    for _ in 0..loops {
+        let _ = coarsetime::Instant::now();
+    }
+    now.elapsed()
+}
+
+#[expect(clippy::explicit_counter_loop)]
+fn primitive(loops: usize) -> Duration {
+    let mut cnt = 0usize;
+    let now = Instant::now();
+    for _ in 0..loops {
+        cnt += 1;
+        let _ = cnt;
+    }
+    now.elapsed()
+}
+
+fn atomic(loops: usize, atomic: Arc<AtomicUsize>) -> Duration {
+    let now = Instant::now();
+    for _ in 0..loops {
+        let _ = atomic.fetch_add(1, Ordering::Relaxed);
+    }
+    now.elapsed()
+}
+
+fn atomic_skip(loops: usize, atomic: Arc<AtomicUsize>, skip: usize) -> Duration {
+    let mut cnt = 0usize;
+    let now = Instant::now();
+    for _ in 0..loops {
+        cnt += 1;
+        let _ = cnt;
+        if cnt % skip == 0 {
+            let _ = atomic.fetch_add(skip, Ordering::Relaxed);
+        } else {
+            let _ = atomic.load(Ordering::Relaxed);
+        }
+    }
+    now.elapsed()
+}
+
+fn sequencer(loops: usize, step: Sequence, lag_amp: Sequence) -> Duration {
+    let sequencer = match (step, lag_amp) {
+        (64, 8) => &SEQUENCER_64_8,
+        (64, 16) => &SEQUENCER_64_16,
+        (64, 32) => &SEQUENCER_64_32,
+        (128, 8) => &SEQUENCER_128_8,
+        (128, 16) => &SEQUENCER_128_16,
+        (128, 32) => &SEQUENCER_128_32,
+        _ => unimplemented!(),
+    };
+    let now = Instant::now();
+    for _ in 0..loops {
+        let _ = sequencer.with(|s| s.borrow_mut().alloc());
+    }
+    now.elapsed()
+}
+
+fn benchmark<F>(name: &str, threads: usize, loops: usize, f: F)
+where
+    F: Fn() -> Duration + Clone + Send + 'static,
+{
+    let handles = (0..threads)
+        .map(|_| std::thread::spawn(black_box(f.clone())))
+        .collect_vec();
+    let mut dur = Duration::from_nanos(0);
+    for handle in handles {
+        dur += handle.join().unwrap();
+    }
+    println!(
+        "{:20} {} threads {} loops: {:?} per iter",
+        name,
+        threads,
+        loops,
+        Duration::from_nanos((dur.as_nanos() / threads as u128 / loops as u128) as u64)
+    );
+}
+
+fn main() {
+    for (threads, loops) in [
+        (1, 10_000_000),
+        (4, 10_000_000),
+        (8, 10_000_000),
+        (16, 10_000_000),
+        (32, 10_000_000),
+    ] {
+        println!();
+
+        benchmark("primitive", threads, loops, move || primitive(loops));
+
+        let a = Arc::new(AtomicUsize::new(0));
+        benchmark("atomic", threads, loops, move || atomic(loops, a.clone()));
+
+        let a = Arc::new(AtomicUsize::new(0));
+        benchmark("atomic skip 8", threads, loops, move || {
+            atomic_skip(loops, a.clone(), 8)
+        });
+
+        let a = Arc::new(AtomicUsize::new(0));
+        benchmark("atomic skip 16", threads, loops, move || {
+            atomic_skip(loops, a.clone(), 16)
+        });
+
+        let a = Arc::new(AtomicUsize::new(0));
+        benchmark("atomic skip 32", threads, loops, move || {
+            atomic_skip(loops, a.clone(), 32)
+        });
+
+        let a = Arc::new(AtomicUsize::new(0));
+        benchmark("atomic skip 64", threads, loops, move || {
+            atomic_skip(loops, a.clone(), 64)
+        });
+
+        benchmark("sequencer(64,8)", threads, loops, move || {
+            sequencer(loops, 64, 8)
+        });
+        benchmark("sequencer(64,16)", threads, loops, move || {
+            sequencer(loops, 64, 16)
+        });
+        benchmark("sequencer(64,32)", threads, loops, move || {
+            sequencer(loops, 64, 32)
+        });
+        benchmark("sequencer(128,8)", threads, loops, move || {
+            sequencer(loops, 128, 8)
+        });
+        benchmark("sequencer(128,16)", threads, loops, move || {
+            sequencer(loops, 128, 16)
+        });
+        benchmark("sequencer(128,32)", threads, loops, move || {
+            sequencer(loops, 128, 32)
+        });
+
+        benchmark("coarse", threads, loops, move || coarse(loops));
+    }
+}