1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
|
#pragma clang system_header
// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements. See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership. The ASF licenses this file
// to you 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.
#include <algorithm>
#include <cstdint>
#include <string>
#include "benchmark/benchmark.h"
#include "contrib/libs/apache/arrow_next/cpp/src/arrow/memory_pool.h"
#include "contrib/libs/apache/arrow_next/cpp/src/arrow/type_fwd.h"
#include "contrib/libs/apache/arrow_next/cpp/src/arrow/util/cpu_info.h"
#include "contrib/libs/apache/arrow_next/cpp/src/arrow/util/logging.h" // IWYU pragma: keep
namespace arrow20 {
// Benchmark changed its parameter type between releases from
// int to int64_t. As it doesn't have version macros, we need
// to apply C++ template magic.
template <typename Func>
struct BenchmarkArgsType;
// Pattern matching that extracts the vector element type of Benchmark::Args()
template <typename Values>
struct BenchmarkArgsType<benchmark::internal::Benchmark* (
benchmark::internal::Benchmark::*)(const std::vector<Values>&)> {
using type = Values;
};
using ArgsType =
typename BenchmarkArgsType<decltype(&benchmark::internal::Benchmark::Args)>::type;
using internal::CpuInfo;
static const CpuInfo* cpu_info = CpuInfo::GetInstance();
static const int64_t kL1Size = cpu_info->CacheSize(CpuInfo::CacheLevel::L1);
static const int64_t kL2Size = cpu_info->CacheSize(CpuInfo::CacheLevel::L2);
static const int64_t kL3Size = cpu_info->CacheSize(CpuInfo::CacheLevel::L3);
static const int64_t kCantFitInL3Size = kL3Size * 4;
static const std::vector<int64_t> kMemorySizes = {kL1Size, kL2Size, kL3Size,
kCantFitInL3Size};
// 0 is treated as "no nulls"
static const std::vector<ArgsType> kInverseNullProportions = {10000, 100, 10, 2, 1, 0};
struct GenericItemsArgs {
// number of items processed per iteration
const int64_t size;
// proportion of nulls in generated arrays
double null_proportion;
explicit GenericItemsArgs(benchmark::State& state)
: size(state.range(0)), state_(state) {
if (state.range(1) == 0) {
this->null_proportion = 0.0;
} else {
this->null_proportion = std::min(1., 1. / static_cast<double>(state.range(1)));
}
}
~GenericItemsArgs() {
state_.counters["size"] = static_cast<double>(size);
state_.counters["null_percent"] = null_proportion * 100;
state_.SetItemsProcessed(state_.iterations() * size);
}
private:
benchmark::State& state_;
};
void BenchmarkSetArgsWithSizes(benchmark::internal::Benchmark* bench,
const std::vector<int64_t>& sizes = kMemorySizes) {
bench->Unit(benchmark::kMicrosecond);
for (const auto size : sizes) {
for (const auto inverse_null_proportion : kInverseNullProportions) {
bench->Args({static_cast<ArgsType>(size), inverse_null_proportion});
}
}
}
void BenchmarkSetArgs(benchmark::internal::Benchmark* bench) {
BenchmarkSetArgsWithSizes(bench, kMemorySizes);
}
void RegressionSetArgs(benchmark::internal::Benchmark* bench) {
// Regression do not need to account for cache hierarchy, thus optimize for
// the best case.
BenchmarkSetArgsWithSizes(bench, {kL1Size});
}
// RAII struct to handle some of the boilerplate in regression benchmarks
struct RegressionArgs {
// size of memory tested (per iteration) in bytes
int64_t size;
// proportion of nulls in generated arrays
double null_proportion;
// If size_is_bytes is true, then it's a number of bytes, otherwise it's the
// number of items processed (for reporting)
explicit RegressionArgs(benchmark::State& state, bool size_is_bytes = true)
: size(state.range(0)), state_(state), size_is_bytes_(size_is_bytes) {
if (state.range(1) == 0) {
this->null_proportion = 0.0;
} else {
this->null_proportion = std::min(1., 1. / static_cast<double>(state.range(1)));
}
}
~RegressionArgs() {
state_.counters["size"] = static_cast<double>(size);
state_.counters["null_percent"] = null_proportion * 100;
if (size_is_bytes_) {
state_.SetBytesProcessed(state_.iterations() * size);
} else {
state_.SetItemsProcessed(state_.iterations() * size);
}
}
private:
benchmark::State& state_;
bool size_is_bytes_;
};
class MemoryPoolMemoryManager : public benchmark::MemoryManager {
void Start() override {
memory_pool = std::make_shared<ProxyMemoryPool>(default_memory_pool());
MemoryPool* default_pool = default_memory_pool();
global_allocations_start = default_pool->num_allocations();
}
// BENCHMARK_DONT_OPTIMIZE is used here to detect Google Benchmark
// 1.8.0. We can remove this Stop(Result*) when we require Google
// Benchmark 1.8.0 or later.
#ifndef BENCHMARK_DONT_OPTIMIZE
void Stop(Result* result) override { Stop(*result); }
#endif
void Stop(benchmark::MemoryManager::Result& result) override {
// If num_allocations is still zero, we assume that the memory pool wasn't passed down
// so we should record them.
MemoryPool* default_pool = default_memory_pool();
int64_t new_default_allocations =
default_pool->num_allocations() - global_allocations_start;
// Only record metrics if (1) there were allocations and (2) we
// recorded at least one.
if (new_default_allocations > 0 && memory_pool->num_allocations() > 0) {
if (new_default_allocations > memory_pool->num_allocations()) {
// If we missed some, let's report that.
int64_t missed_allocations =
new_default_allocations - memory_pool->num_allocations();
ARROW_LOG(WARNING) << "BenchmarkMemoryTracker recorded some allocations "
<< "for a benchmark, but missed " << missed_allocations
<< " allocations.\n";
}
result.max_bytes_used = memory_pool->max_memory();
result.total_allocated_bytes = memory_pool->total_bytes_allocated();
result.num_allocs = memory_pool->num_allocations();
}
}
public:
std::shared_ptr<::arrow20::ProxyMemoryPool> memory_pool;
protected:
int64_t global_allocations_start;
};
/// \brief Track memory pool allocations in benchmarks.
///
/// Instantiate as a global variable to register the hooks into Google Benchmark
/// to collect memory metrics. Before each benchmark, a new ProxyMemoryPool is
/// created. It can then be accessed with memory_pool(). Once the benchmark is
/// complete, the hook will record the maximum memory used, the total bytes
/// allocated, and the total number of allocations. If no allocations were seen,
/// (for example, if you forgot to pass down the memory pool), then these metrics
/// will not be saved.
///
/// Since this is used as one global variable, this will not work if multiple
/// benchmarks are run concurrently or for multi-threaded benchmarks (ones
/// that use `->ThreadRange(...)`).
class BenchmarkMemoryTracker {
public:
BenchmarkMemoryTracker() : manager_() { ::benchmark::RegisterMemoryManager(&manager_); }
::arrow20::MemoryPool* memory_pool() const { return manager_.memory_pool.get(); }
protected:
::arrow20::MemoryPoolMemoryManager manager_;
};
} // namespace arrow20
|
