tdlight/tdutils/test/hashset_benchmark.cpp

648 lines
18 KiB
C++

//
// Copyright Aliaksei Levin (levlam@telegram.org), Arseny Smirnov (arseny30@gmail.com) 2014-2022
//
// Distributed under the Boost Software License, Version 1.0. (See accompanying
// file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
//
#include "td/utils/algorithm.h"
#include "td/utils/common.h"
#include "td/utils/FlatHashMap.h"
#include "td/utils/FlatHashMapChunks.h"
#include "td/utils/FlatHashTable.h"
#include "td/utils/format.h"
#include "td/utils/Hash.h"
#include "td/utils/logging.h"
#include "td/utils/MapNode.h"
#include "td/utils/Random.h"
#include "td/utils/Slice.h"
#include "td/utils/Span.h"
#include "td/utils/StringBuilder.h"
#include "td/utils/Time.h"
#include "td/utils/VectorQueue.h"
#ifdef SCOPE_EXIT
#undef SCOPE_EXIT
#endif
#include <absl/container/flat_hash_map.h>
#include <absl/hash/hash.h>
#include <algorithm>
#include <benchmark/benchmark.h>
#include <folly/container/F14Map.h>
#include <functional>
#include <map>
#include <random>
#include <unordered_map>
#include <utility>
template <class TableT>
static void reserve(TableT &table, std::size_t size) {
table.reserve(size);
}
template <class A, class B>
static void reserve(std::map<A, B> &table, std::size_t size) {
}
template <class KeyT, class ValueT>
class NoOpTable {
public:
using key_type = KeyT;
using value_type = std::pair<const KeyT, ValueT>;
template <class It>
NoOpTable(It begin, It end) {
}
ValueT &operator[](const KeyT &) const {
static ValueT dummy;
return dummy;
}
KeyT find(const KeyT &key) const {
return key;
}
};
template <class KeyT, class ValueT>
class VectorTable {
public:
using key_type = KeyT;
using value_type = std::pair<const KeyT, ValueT>;
template <class It>
VectorTable(It begin, It end) : table_(begin, end) {
}
ValueT &operator[](const KeyT &needle) {
auto it = find(needle);
if (it == table_.end()) {
table_.emplace_back(needle, ValueT{});
return table_.back().second;
}
return it->second;
}
auto find(const KeyT &needle) {
return std::find_if(table_.begin(), table_.end(), [&](auto &key) { return key.first == needle; });
}
private:
using KeyValue = value_type;
td::vector<KeyValue> table_;
};
template <class KeyT, class ValueT>
class SortedVectorTable {
public:
using key_type = KeyT;
using value_type = std::pair<KeyT, ValueT>;
template <class It>
SortedVectorTable(It begin, It end) : table_(begin, end) {
std::sort(table_.begin(), table_.end());
}
ValueT &operator[](const KeyT &needle) {
auto it = std::lower_bound(table_.begin(), table_.end(), needle,
[](const auto &l, const auto &r) { return l.first < r; });
if (it == table_.end() || it->first != needle) {
it = table_.insert(it, {needle, ValueT{}});
}
return it->second;
}
auto find(const KeyT &needle) {
auto it = std::lower_bound(table_.begin(), table_.end(), needle,
[](const auto &l, const auto &r) { return l.first < r; });
if (it != table_.end() && it->first == needle) {
return it;
}
return table_.end();
}
private:
using KeyValue = value_type;
td::vector<KeyValue> table_;
};
template <class KeyT, class ValueT, class HashT = td::Hash<KeyT>>
class SimpleHashTable {
public:
using key_type = KeyT;
using value_type = std::pair<KeyT, ValueT>;
template <class It>
SimpleHashTable(It begin, It end) {
nodes_.resize((end - begin) * 2);
for (; begin != end; ++begin) {
insert(begin->first, begin->second);
}
}
ValueT &operator[](const KeyT &needle) {
UNREACHABLE();
}
ValueT *find(const KeyT &needle) {
auto hash = HashT()(needle);
std::size_t i = hash % nodes_.size();
while (true) {
if (nodes_[i].key == needle) {
return &nodes_[i].value;
}
if (nodes_[i].hash == 0) {
return nullptr;
}
i++;
if (i == nodes_.size()) {
i = 0;
}
}
}
private:
using KeyValue = value_type;
struct Node {
std::size_t hash{0};
KeyT key;
ValueT value;
};
td::vector<Node> nodes_;
void insert(KeyT key, ValueT value) {
auto hash = HashT()(key);
std::size_t i = hash % nodes_.size();
while (true) {
if (nodes_[i].hash == 0 || (nodes_[i].hash == hash && nodes_[i].key == key)) {
nodes_[i].value = value;
nodes_[i].key = key;
nodes_[i].hash = hash;
return;
}
i++;
if (i == nodes_.size()) {
i = 0;
}
}
}
};
template <typename TableT>
static void BM_Get(benchmark::State &state) {
std::size_t n = state.range(0);
constexpr std::size_t BATCH_SIZE = 1024;
td::Random::Xorshift128plus rnd(123);
using Key = typename TableT::key_type;
using Value = typename TableT::value_type::second_type;
using KeyValue = std::pair<Key, Value>;
td::vector<KeyValue> data;
td::vector<Key> keys;
TableT table;
for (std::size_t i = 0; i < n; i++) {
auto key = rnd();
auto value = rnd();
data.emplace_back(key, value);
table.emplace(key, value);
keys.push_back(key);
}
std::size_t key_i = 0;
td::random_shuffle(td::as_mutable_span(keys), rnd);
auto next_key = [&] {
key_i++;
if (key_i == data.size()) {
key_i = 0;
}
return keys[key_i];
};
while (state.KeepRunningBatch(BATCH_SIZE)) {
for (std::size_t i = 0; i < BATCH_SIZE; i++) {
benchmark::DoNotOptimize(table.find(next_key()));
}
}
}
template <typename TableT>
static void BM_find_same(benchmark::State &state) {
td::Random::Xorshift128plus rnd(123);
TableT table;
constexpr std::size_t N = 100000;
constexpr std::size_t BATCH_SIZE = 1024;
reserve(table, N);
for (std::size_t i = 0; i < N; i++) {
table.emplace(rnd(), i);
}
auto key = td::Random::secure_uint64();
table[key] = 123;
while (state.KeepRunningBatch(BATCH_SIZE)) {
for (std::size_t i = 0; i < BATCH_SIZE; i++) {
benchmark::DoNotOptimize(table.find(key));
}
}
}
template <typename TableT>
static void BM_emplace_same(benchmark::State &state) {
td::Random::Xorshift128plus rnd(123);
TableT table;
constexpr std::size_t N = 100000;
constexpr std::size_t BATCH_SIZE = 1024;
reserve(table, N);
for (std::size_t i = 0; i < N; i++) {
table.emplace(rnd(), i);
}
auto key = 123743;
table[key] = 123;
while (state.KeepRunningBatch(BATCH_SIZE)) {
for (std::size_t i = 0; i < BATCH_SIZE; i++) {
benchmark::DoNotOptimize(table.emplace(key + (i & 15) * 100, 43784932));
}
}
}
template <typename TableT>
static void BM_emplace_string(benchmark::State &state) {
td::Random::Xorshift128plus rnd(123);
TableT table;
constexpr std::size_t N = 100000;
constexpr std::size_t BATCH_SIZE = 1024;
reserve(table, N);
for (std::size_t i = 0; i < N; i++) {
table.emplace(td::to_string(rnd()), i);
}
table["0"] = 123;
td::vector<td::string> strings;
for (std::size_t i = 0; i < 16; i++) {
strings.emplace_back(1, static_cast<char>('0' + i));
}
while (state.KeepRunningBatch(BATCH_SIZE)) {
for (std::size_t i = 0; i < BATCH_SIZE; i++) {
benchmark::DoNotOptimize(table.emplace(strings[i & 15], 43784932));
}
}
}
namespace td {
template <class K, class V, class FunctT>
static void table_remove_if(absl::flat_hash_map<K, V> &table, FunctT &&func) {
for (auto it = table.begin(); it != table.end();) {
if (func(*it)) {
auto copy = it;
++it;
table.erase(copy);
} else {
++it;
}
}
}
} // namespace td
template <typename TableT>
static void BM_remove_if(benchmark::State &state) {
constexpr std::size_t N = 100000;
constexpr std::size_t BATCH_SIZE = N;
TableT table;
reserve(table, N);
while (state.KeepRunningBatch(BATCH_SIZE)) {
state.PauseTiming();
td::Random::Xorshift128plus rnd(123);
for (std::size_t i = 0; i < N; i++) {
table.emplace(rnd(), i);
}
state.ResumeTiming();
td::table_remove_if(table, [](auto &it) { return it.second % 2 == 0; });
}
}
template <typename TableT>
static void BM_erase_all_with_begin(benchmark::State &state) {
constexpr std::size_t N = 100000;
constexpr std::size_t BATCH_SIZE = N;
TableT table;
td::Random::Xorshift128plus rnd(123);
while (state.KeepRunningBatch(BATCH_SIZE)) {
for (std::size_t i = 0; i < BATCH_SIZE; i++) {
table.emplace(rnd() + 1, i);
}
while (!table.empty()) {
table.erase(table.begin());
}
}
}
template <typename TableT>
static void BM_cache(benchmark::State &state) {
constexpr std::size_t N = 1000;
constexpr std::size_t BATCH_SIZE = 1000000;
TableT table;
td::Random::Xorshift128plus rnd(123);
td::VectorQueue<td::uint64> keys;
while (state.KeepRunningBatch(BATCH_SIZE)) {
for (std::size_t i = 0; i < BATCH_SIZE; i++) {
auto key = rnd() + 1;
keys.push(key);
table.emplace(key, i);
if (table.size() > N) {
table.erase(keys.pop());
}
}
}
}
template <typename TableT>
static void BM_cache2(benchmark::State &state) {
constexpr std::size_t N = 1000;
constexpr std::size_t BATCH_SIZE = 1000000;
TableT table;
td::Random::Xorshift128plus rnd(123);
td::VectorQueue<td::uint64> keys;
while (state.KeepRunningBatch(BATCH_SIZE)) {
for (std::size_t i = 0; i < BATCH_SIZE; i++) {
auto key = rnd() + 1;
keys.push(key);
table.emplace(key, i);
if (table.size() > N) {
table.erase(keys.pop_rand(rnd));
}
}
}
}
template <typename TableT>
static void BM_cache3(benchmark::State &state) {
std::size_t N = state.range(0);
constexpr std::size_t BATCH_SIZE = 1000000;
TableT table;
td::Random::Xorshift128plus rnd(123);
td::VectorQueue<td::uint64> keys;
std::size_t step = 20;
while (state.KeepRunningBatch(BATCH_SIZE)) {
for (std::size_t i = 0; i < BATCH_SIZE; i += step) {
auto key = rnd() + 1;
keys.push(key);
table.emplace(key, i);
for (std::size_t j = 1; j < step; j++) {
auto key_to_find = keys.data()[rnd() % keys.size()];
benchmark::DoNotOptimize(table.find(key_to_find));
}
if (table.size() > N) {
table.erase(keys.pop_rand(rnd));
}
}
}
}
template <typename TableT>
static void BM_remove_if_slow(benchmark::State &state) {
constexpr std::size_t N = 5000;
constexpr std::size_t BATCH_SIZE = 500000;
TableT table;
td::Random::Xorshift128plus rnd(123);
for (std::size_t i = 0; i < N; i++) {
table.emplace(rnd() + 1, i);
}
auto first_key = table.begin()->first;
{
std::size_t cnt = 0;
td::table_remove_if(table, [&cnt, n = N](auto &) {
cnt += 2;
return cnt <= n;
});
}
while (state.KeepRunningBatch(BATCH_SIZE)) {
for (std::size_t i = 0; i < BATCH_SIZE; i++) {
table.emplace(first_key, i);
table.erase(first_key);
}
}
}
template <typename TableT>
static void BM_remove_if_slow_old(benchmark::State &state) {
constexpr std::size_t N = 100000;
constexpr std::size_t BATCH_SIZE = 5000000;
TableT table;
while (state.KeepRunningBatch(BATCH_SIZE)) {
td::Random::Xorshift128plus rnd(123);
for (std::size_t i = 0; i < BATCH_SIZE; i++) {
table.emplace(rnd() + 1, i);
if (table.size() > N) {
std::size_t cnt = 0;
td::table_remove_if(table, [&cnt, n = N](auto &) {
cnt += 2;
return cnt <= n;
});
}
}
}
}
template <typename TableT>
static void benchmark_create(td::Slice name) {
td::Random::Xorshift128plus rnd(123);
{
constexpr std::size_t N = 10000000;
TableT table;
reserve(table, N);
auto start = td::Timestamp::now();
for (std::size_t i = 0; i < N; i++) {
table.emplace(rnd(), i);
}
auto end = td::Timestamp::now();
LOG(INFO) << name << ": create " << N << " elements: " << td::format::as_time(end.at() - start.at());
double res = 0;
td::vector<std::pair<std::size_t, td::format::Time>> pauses;
for (std::size_t i = 0; i < N; i++) {
auto emplace_start = td::Timestamp::now();
table.emplace(rnd(), i);
auto emplace_end = td::Timestamp::now();
auto pause = emplace_end.at() - emplace_start.at();
res = td::max(pause, res);
if (pause > 0.001) {
pauses.emplace_back(i, td::format::as_time(pause));
}
}
LOG(INFO) << name << ": create another " << N << " elements, max pause = " << td::format::as_time(res) << " "
<< pauses;
}
}
struct CacheMissNode {
td::uint32 data{};
char padding[64 - sizeof(data)];
};
class IterateFast {
public:
static td::uint32 iterate(CacheMissNode *ptr, std::size_t max_shift) {
td::uint32 res = 1;
for (std::size_t i = 0; i < max_shift; i++) {
if (ptr[i].data % max_shift != 0) {
res *= ptr[i].data;
} else {
res /= ptr[i].data;
}
}
return res;
}
};
class IterateSlow {
public:
static td::uint32 iterate(CacheMissNode *ptr, std::size_t max_shift) {
td::uint32 res = 1;
for (std::size_t i = 0;; i++) {
if (ptr[i].data % max_shift != 0) {
res *= ptr[i].data;
} else {
break;
}
}
return res;
}
};
template <class F>
static void BM_cache_miss(benchmark::State &state) {
td::uint32 max_shift = state.range(0);
bool flag = state.range(1);
std::random_device rd;
std::mt19937 rnd(rd());
int N = 50000000;
td::vector<CacheMissNode> nodes(N);
td::uint32 i = 0;
for (auto &node : nodes) {
if (flag) {
node.data = i++ % max_shift;
} else {
node.data = rnd();
}
}
td::vector<int> positions(N);
std::uniform_int_distribution<td::uint32> rnd_pos(0, N - 1000);
for (auto &pos : positions) {
pos = rnd_pos(rnd);
if (flag) {
pos = pos / max_shift * max_shift + 1;
}
}
while (state.KeepRunningBatch(positions.size())) {
for (const auto pos : positions) {
auto *ptr = &nodes[pos];
auto res = F::iterate(ptr, max_shift);
benchmark::DoNotOptimize(res);
}
}
}
static td::uint64 equal_mask_slow(td::uint8 *bytes, td::uint8 needle) {
td::uint64 mask = 0;
for (int i = 0; i < 16; i++) {
mask |= (bytes[i] == needle) << i;
}
return mask;
}
template <class MaskT>
static void BM_mask(benchmark::State &state) {
std::size_t BATCH_SIZE = 1024;
td::vector<td::uint8> bytes(BATCH_SIZE + 16);
for (auto &b : bytes) {
b = static_cast<td::uint8>(td::Random::fast(0, 17));
}
while (state.KeepRunningBatch(BATCH_SIZE)) {
for (std::size_t i = 0; i < BATCH_SIZE; i++) {
benchmark::DoNotOptimize(MaskT::equal_mask(bytes.data() + i, 17));
}
}
}
BENCHMARK_TEMPLATE(BM_mask, td::MaskPortable);
#ifdef __aarch64__
BENCHMARK_TEMPLATE(BM_mask, td::MaskNeonFolly);
BENCHMARK_TEMPLATE(BM_mask, td::MaskNeon);
#endif
#if TD_SSE2
BENCHMARK_TEMPLATE(BM_mask, td::MaskSse2);
#endif
template <class KeyT, class ValueT, class HashT = std::hash<KeyT>, class EqT = std::equal_to<KeyT>>
using FlatHashMapImpl = td::FlatHashTable<td::MapNode<KeyT, ValueT>, HashT, EqT>;
#define FOR_EACH_TABLE(F) \
F(FlatHashMapImpl) \
F(td::FlatHashMapChunks) \
F(folly::F14FastMap) \
F(absl::flat_hash_map) \
F(std::unordered_map) \
F(std::map)
//BENCHMARK(BM_cache_miss<IterateSlow>)->Ranges({{1, 16}, {0, 1}});
//BENCHMARK(BM_cache_miss<IterateFast>)->Ranges({{1, 16}, {0, 1}});
//BENCHMARK_TEMPLATE(BM_Get, VectorTable<td::uint64, td::uint64>)->Range(1, 1 << 26);
//BENCHMARK_TEMPLATE(BM_Get, SortedVectorTable<td::uint64, td::uint64>)->Range(1, 1 << 26);
//BENCHMARK_TEMPLATE(BM_Get, NoOpTable<td::uint64, td::uint64>)->Range(1, 1 << 26);
#define REGISTER_GET_BENCHMARK(HT) BENCHMARK_TEMPLATE(BM_Get, HT<td::uint64, td::uint64>)->Range(1, 1 << 23);
#define REGISTER_FIND_BENCHMARK(HT) \
BENCHMARK_TEMPLATE(BM_find_same, HT<td::uint64, td::uint64>) \
->ComputeStatistics("max", [](const td::vector<double> &v) { return *std::max_element(v.begin(), v.end()); }) \
->ComputeStatistics("min", [](const td::vector<double> &v) { return *std::min_element(v.begin(), v.end()); }) \
->Repetitions(20) \
->DisplayAggregatesOnly(true);
#define REGISTER_REMOVE_IF_BENCHMARK(HT) BENCHMARK_TEMPLATE(BM_remove_if, HT<td::uint64, td::uint64>);
#define REGISTER_EMPLACE_BENCHMARK(HT) BENCHMARK_TEMPLATE(BM_emplace_same, HT<td::uint64, td::uint64>);
#define REGISTER_EMPLACE_STRING_BENCHMARK(HT) BENCHMARK_TEMPLATE(BM_emplace_string, HT<td::string, td::uint64>);
#define REGISTER_CACHE_BENCHMARK(HT) BENCHMARK_TEMPLATE(BM_cache, HT<td::uint64, td::uint64>);
#define REGISTER_CACHE2_BENCHMARK(HT) BENCHMARK_TEMPLATE(BM_cache2, HT<td::uint64, td::uint64>);
#define REGISTER_CACHE3_BENCHMARK(HT) BENCHMARK_TEMPLATE(BM_cache3, HT<td::uint64, td::uint64>)->Range(1, 1 << 23);
#define REGISTER_ERASE_ALL_BENCHMARK(HT) BENCHMARK_TEMPLATE(BM_erase_all_with_begin, HT<td::uint64, td::uint64>);
#define REGISTER_REMOVE_IF_SLOW_BENCHMARK(HT) BENCHMARK_TEMPLATE(BM_remove_if_slow, HT<td::uint64, td::uint64>);
#define REGISTER_REMOVE_IF_SLOW_OLD_BENCHMARK(HT) BENCHMARK_TEMPLATE(BM_remove_if_slow_old, HT<td::uint64, td::uint64>);
FOR_EACH_TABLE(REGISTER_GET_BENCHMARK)
FOR_EACH_TABLE(REGISTER_CACHE3_BENCHMARK)
FOR_EACH_TABLE(REGISTER_CACHE2_BENCHMARK)
FOR_EACH_TABLE(REGISTER_CACHE_BENCHMARK)
FOR_EACH_TABLE(REGISTER_REMOVE_IF_BENCHMARK)
FOR_EACH_TABLE(REGISTER_EMPLACE_BENCHMARK)
FOR_EACH_TABLE(REGISTER_EMPLACE_STRING_BENCHMARK)
FOR_EACH_TABLE(REGISTER_ERASE_ALL_BENCHMARK)
FOR_EACH_TABLE(REGISTER_FIND_BENCHMARK)
FOR_EACH_TABLE(REGISTER_REMOVE_IF_SLOW_OLD_BENCHMARK)
FOR_EACH_TABLE(REGISTER_REMOVE_IF_SLOW_BENCHMARK)
#define RUN_CREATE_BENCHMARK(HT) benchmark_create<HT<td::uint64, td::uint64>>(#HT);
int main(int argc, char **argv) {
// FOR_EACH_TABLE(RUN_CREATE_BENCHMARK);
benchmark::Initialize(&argc, argv);
benchmark::RunSpecifiedBenchmarks();
benchmark::Shutdown();
}