Improve parallelization with fine-grained locks

This commit is contained in:
Dirkjan Ochtman 2021-01-21 13:45:39 +01:00
parent f388fd0a46
commit 58bb0f315a
3 changed files with 132 additions and 122 deletions

View File

@ -9,6 +9,7 @@ edition = "2018"
indicatif = { version = "0.15", optional = true }
num_cpus = "1.13"
ordered-float = "2.0"
parking_lot = "0.11"
rand = { version = "0.8", features = ["small_rng"] }
rayon = "1.5"
serde = { version = "1.0.118", features = ["derive"], optional = true }

View File

@ -1,13 +1,16 @@
use std::cmp::{max, min, Ordering, Reverse};
use std::cmp::{max, Ordering, Reverse};
use std::collections::BinaryHeap;
use std::collections::HashSet;
#[cfg(feature = "indicatif")]
use std::sync::atomic::{self, AtomicUsize};
#[cfg(feature = "indicatif")]
use indicatif::ProgressBar;
use ordered_float::OrderedFloat;
use parking_lot::{Mutex, RwLock};
use rand::rngs::SmallRng;
use rand::SeedableRng;
use rayon::iter::{IntoParallelRefMutIterator, ParallelIterator};
use rayon::iter::{IntoParallelIterator, ParallelIterator};
#[cfg(feature = "serde")]
use serde::{Deserialize, Serialize};
@ -125,6 +128,7 @@ where
let ef_search = builder.ef_search.unwrap_or(100);
let ef_construction = builder.ef_construction.unwrap_or(100);
let ml = builder.ml.unwrap_or_else(|| (M as f32).ln());
let heuristic = builder.heuristic;
let mut rng = match builder.seed {
Some(seed) => SmallRng::seed_from_u64(seed),
None => SmallRng::from_entropy(),
@ -136,6 +140,7 @@ where
if let Some(bar) = &progress {
bar.set_draw_delta(1_000);
bar.set_length(points.len() as u64);
bar.set_message("Build index (preparation)");
}
if points.is_empty() {
@ -220,88 +225,67 @@ where
// Insert the first point so that we have an enter point to start searches with.
let mut layers = vec![vec![]; top.0];
let mut zero = Vec::with_capacity(points.len());
zero.push(ZeroNode::default());
let zero = points
.iter()
.map(|_| RwLock::new(ZeroNode::default()))
.collect::<Vec<_>>();
let mut insertion = Search {
ef: ef_construction,
visited: Visited::with_capacity(points.len()),
..Default::default()
};
let mut pool = SearchPool {
pool: Vec::new(),
len: points.len(),
};
let mut batch = Vec::new();
let mut done = Vec::new();
let max_batch_len = num_cpus::get() * 4;
for (layer, mut range) in ranges {
let pool = SearchPool::new(points.len());
#[cfg(feature = "indicatif")]
let done = AtomicUsize::new(0);
for (layer, range) in ranges {
let num = if layer.0 > 0 { M } else { M * 2 };
#[cfg(feature = "indicatif")]
if let Some(bar) = &progress {
bar.set_message(&format!("Building index (layer {})", layer.0));
}
while range.start < range.end {
let len = min(range.len(), max_batch_len);
batch.clear();
batch.extend(
nodes[range.start..(range.start + len)]
.iter()
.map(|&(_, pid)| (pid, pool.pop())),
);
nodes[range].into_par_iter().for_each(|(_, pid)| {
let (mut search, mut insertion) = pool.pop();
let point = &points.as_slice()[*pid];
search.reset();
search.push(PointId(0), point, &points);
batch.par_iter_mut().for_each(|(pid, search)| {
let point = &points[*pid];
search.push(PointId(0), point, &points);
for cur in top.descend() {
search.ef = if cur <= layer { ef_construction } else { 1 };
match cur > layer {
true => {
search.search(point, layers[cur.0 - 1].as_slice(), &points, num);
search.cull();
}
false => {
search.search(point, zero.as_slice(), &points, num);
break;
}
for cur in top.descend() {
search.ef = if cur <= layer { ef_construction } else { 1 };
match cur > layer {
true => {
search.search(point, layers[cur.0 - 1].as_slice(), &points, num);
search.cull();
}
false => {
search.search(point, zero.as_slice(), &points, num);
break;
}
}
});
done.clear();
for (pid, mut search) in batch.drain(..) {
for added in done.iter().copied() {
search.push(added, &points[pid], &points);
}
insert(
pid,
&mut insertion,
&mut search,
&mut zero,
&points,
&builder.heuristic,
);
done.push(pid);
pool.push(search);
}
insertion.ef = ef_construction;
insert(
*pid,
&mut insertion,
&mut search,
&zero,
&points,
&heuristic,
);
#[cfg(feature = "indicatif")]
if let Some(bar) = &progress {
bar.inc(done.len() as u64);
let value = done.fetch_add(1, atomic::Ordering::Relaxed);
if value % 1000 == 0 {
bar.set_position(value as u64);
}
}
range.start += len;
}
pool.push((search, insertion));
});
// For layers above the zero layer, make a copy of the current state of the zero layer
// with `nearest` truncated to `M` elements.
if layer.0 > 0 {
let mut upper = Vec::with_capacity(zero.len());
upper.extend(zero.iter().map(UpperNode::from_zero));
upper.extend(zero.iter().map(|zero| UpperNode::from_zero(&zero.read())));
layers[layer.0 - 1] = upper;
}
}
@ -314,7 +298,7 @@ where
(
Self {
ef_search,
zero,
zero: zero.into_iter().map(|node| node.into_inner()).collect(),
points,
layers,
},
@ -382,14 +366,14 @@ fn insert<P: Point>(
new: PointId,
insertion: &mut Search,
search: &mut Search,
layer: &mut Vec<ZeroNode>,
layer: &[RwLock<ZeroNode>],
points: &[P],
heuristic: &Option<Heuristic>,
) {
layer.push(ZeroNode::default());
let mut node = layer[new].write();
let found = match heuristic {
None => search.select_simple(M * 2),
Some(heuristic) => search.select_heuristic(&points[new], &layer, points, *heuristic),
Some(heuristic) => search.select_heuristic(&points[new], layer, points, *heuristic),
};
// Just make sure the candidates are all unique
@ -404,19 +388,22 @@ fn insert<P: Point>(
if let Some(heuristic) = heuristic {
let found = insertion.add_neighbor_heuristic(
new,
layer.as_slice().nearest_iter(pid),
layer.nearest_iter(pid),
layer,
&points[pid],
points,
*heuristic,
);
layer[pid].rewrite(found.iter().map(|candidate| candidate.pid));
layer[new].set(i, pid);
layer[pid]
.write()
.rewrite(found.iter().map(|candidate| candidate.pid));
node.set(i, pid);
} else {
// Find the correct index to insert at to keep the neighbor's neighbors sorted
let old = &points[pid];
let idx = layer[pid]
.read()
.binary_search_by(|third| {
// `third` here is one of the neighbors of the new node's neighbor.
let third = match third {
@ -429,30 +416,34 @@ fn insert<P: Point>(
})
.unwrap_or_else(|e| e);
layer[pid].insert(idx, new);
layer[new].set(i, pid);
layer[pid].write().insert(idx, new);
node.set(i, pid);
}
}
}
struct SearchPool {
pool: Vec<Search>,
pool: Mutex<Vec<(Search, Search)>>,
len: usize,
}
impl SearchPool {
fn pop(&mut self) -> Search {
match self.pool.pop() {
Some(mut search) => {
search.reset();
search
}
None => Search::new(self.len),
fn new(len: usize) -> Self {
Self {
pool: Mutex::new(Vec::new()),
len,
}
}
fn push(&mut self, search: Search) {
self.pool.push(search);
fn pop(&self) -> (Search, Search) {
match self.pool.lock().pop() {
Some(res) => res,
None => (Search::new(self.len), Search::new(self.len)),
}
}
fn push(&self, item: (Search, Search)) {
self.pool.lock().push(item);
}
}
@ -516,11 +507,11 @@ impl Search {
}
}
fn add_neighbor_heuristic<P: Point>(
fn add_neighbor_heuristic<L: Layer, P: Point>(
&mut self,
new: PointId,
current: impl Iterator<Item = PointId>,
layer: &[ZeroNode],
layer: L,
point: &P,
points: &[P],
params: Heuristic,
@ -530,16 +521,16 @@ impl Search {
for pid in current {
self.push(pid, point, points);
}
self.select_heuristic(point, &layer, points, params)
self.select_heuristic(point, layer, points, params)
}
/// Heuristically sort and truncate neighbors in `self.nearest`
///
/// Invariant: `self.nearest` must be in sorted (nearest first) order.
fn select_heuristic<P: Point>(
fn select_heuristic<L: Layer, P: Point>(
&mut self,
point: &P,
layer: &[ZeroNode],
layer: L,
points: &[P],
params: Heuristic,
) -> &[Candidate] {

View File

@ -1,7 +1,8 @@
use std::hash::Hash;
use std::ops::{Deref, Index, IndexMut};
use std::ops::{Deref, Index};
use ordered_float::OrderedFloat;
use parking_lot::{MappedRwLockReadGuard, RwLock, RwLockReadGuard};
use rand::rngs::SmallRng;
use rand::Rng;
#[cfg(feature = "serde")]
@ -69,11 +70,11 @@ impl UpperNode {
}
}
impl Layer for &[UpperNode] {
fn nearest_iter(&self, pid: PointId) -> NearestIter<'_> {
NearestIter {
nearest: &self[pid.0 as usize].0,
}
impl<'a> Layer for &'a [UpperNode] {
type Slice = &'a [PointId];
fn nearest_iter(&self, pid: PointId) -> NearestIter<Self::Slice> {
NearestIter::new(&self[pid.0 as usize].0)
}
}
@ -128,32 +129,63 @@ impl Deref for ZeroNode {
}
}
impl Layer for &[ZeroNode] {
fn nearest_iter(&self, pid: PointId) -> NearestIter<'_> {
NearestIter {
nearest: &self[pid.0 as usize].0,
}
impl<'a> Layer for &'a [ZeroNode] {
type Slice = &'a [PointId];
fn nearest_iter(&self, pid: PointId) -> NearestIter<Self::Slice> {
NearestIter::new(&self[pid.0 as usize])
}
}
impl<'a> Layer for &'a [RwLock<ZeroNode>] {
type Slice = MappedRwLockReadGuard<'a, [PointId]>;
fn nearest_iter(&self, pid: PointId) -> NearestIter<Self::Slice> {
NearestIter::new(RwLockReadGuard::map(
self[pid.0 as usize].read(),
Deref::deref,
))
}
}
pub(crate) trait Layer {
fn nearest_iter(&self, pid: PointId) -> NearestIter<'_>;
type Slice: Deref<Target = [PointId]>;
fn nearest_iter(&self, pid: PointId) -> NearestIter<Self::Slice>;
}
pub(crate) struct NearestIter<'a> {
nearest: &'a [PointId],
pub(crate) struct NearestIter<T> {
node: T,
cur: usize,
}
impl<'a> Iterator for NearestIter<'a> {
impl<T> NearestIter<T>
where
T: Deref<Target = [PointId]>,
{
fn new(node: T) -> Self {
Self { node, cur: 0 }
}
}
impl<T> Iterator for NearestIter<T>
where
T: Deref<Target = [PointId]>,
{
type Item = PointId;
fn next(&mut self) -> Option<Self::Item> {
let (&first, rest) = self.nearest.split_first()?;
if !first.is_valid() {
if self.cur >= self.node.len() {
return None;
}
self.nearest = rest;
Some(first)
let item = self.node[self.cur];
if !item.is_valid() {
self.cur = self.node.len();
return None;
}
self.cur += 1;
Some(item)
}
}
@ -230,14 +262,6 @@ impl<P> Index<PointId> for Hnsw<P> {
}
}
impl<P: Point> Index<PointId> for Vec<P> {
type Output = P;
fn index(&self, index: PointId) -> &Self::Output {
&self[index.0 as usize]
}
}
impl<P: Point> Index<PointId> for [P] {
type Output = P;
@ -246,18 +270,12 @@ impl<P: Point> Index<PointId> for [P] {
}
}
impl Index<PointId> for Vec<ZeroNode> {
type Output = ZeroNode;
impl Index<PointId> for [RwLock<ZeroNode>] {
type Output = RwLock<ZeroNode>;
fn index(&self, index: PointId) -> &Self::Output {
&self[index.0 as usize]
}
}
impl IndexMut<PointId> for Vec<ZeroNode> {
fn index_mut(&mut self, index: PointId) -> &mut Self::Output {
&mut self[index.0 as usize]
}
}
pub(crate) const INVALID: PointId = PointId(u32::MAX);