summaryrefslogtreecommitdiff
path: root/src/mongo/unittest/system_resource_canary_bm.cpp
blob: 945b46b7b22e360cfc650c425dbdf19897c0369c (plain)
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
/**
 *    Copyright (C) 2018-present MongoDB, Inc.
 *
 *    This program is free software: you can redistribute it and/or modify
 *    it under the terms of the Server Side Public License, version 1,
 *    as published by MongoDB, Inc.
 *
 *    This program is distributed in the hope that it will be useful,
 *    but WITHOUT ANY WARRANTY; without even the implied warranty of
 *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *    Server Side Public License for more details.
 *
 *    You should have received a copy of the Server Side Public License
 *    along with this program. If not, see
 *    <http://www.mongodb.com/licensing/server-side-public-license>.
 *
 *    As a special exception, the copyright holders give permission to link the
 *    code of portions of this program with the OpenSSL library under certain
 *    conditions as described in each individual source file and distribute
 *    linked combinations including the program with the OpenSSL library. You
 *    must comply with the Server Side Public License in all respects for
 *    all of the code used other than as permitted herein. If you modify file(s)
 *    with this exception, you may extend this exception to your version of the
 *    file(s), but you are not obligated to do so. If you do not wish to do so,
 *    delete this exception statement from your version. If you delete this
 *    exception statement from all source files in the program, then also delete
 *    it in the license file.
 */

#include "mongo/platform/basic.h"

#include <benchmark/benchmark.h>

#include "mongo/util/assert_util.h"
#include "mongo/util/processinfo.h"
#include "mongo/util/time_support.h"

namespace mongo {
namespace {

// This is a trivial test case to sanity check that "benchmark" runs.
// For more information on how benchmarks should be written, please refer to Google Benchmark's
// excellent README: https://github.com/google/benchmark/blob/v1.3.0/README.md
void BM_Empty(benchmark::State& state) {
    for (auto keepRunning : state) {
        // The code inside this for-loop is what's being timed.
        benchmark::DoNotOptimize(state.iterations());
    }
}

// Register two benchmarks, one runs the "BM_Empty" function in a single thread, the other runs a
// copy per CPU core.
BENCHMARK(BM_Empty);
BENCHMARK(BM_Empty)->ThreadPerCpu();

void BM_CpuLoad(benchmark::State& state) {
    for (auto keepRunning : state) {
        uint64_t limit = 10000;
        uint64_t lresult = 0;
        uint64_t x = 100;
        for (uint64_t i = 0; i < limit; i++) {
            benchmark::DoNotOptimize(x *= 13);
        }
        benchmark::DoNotOptimize(lresult = x);
    }
}

// This Benchmark is adapted from the `cpuload` command:
// https://github.com/mongodb/mongo/blob/r3.7.2/src/mongo/db/commands/cpuload.cpp
BENCHMARK(BM_CpuLoad)->Threads(1)->ThreadPerCpu();


void BM_Sleep(benchmark::State& state) {
    for (auto keepRunning : state) {
        sleepmillis(100);
    }
}

BENCHMARK(BM_Sleep)->Threads(1)->ThreadPerCpu();


// Generate a loop with macros.
#define ONE ptrToNextLinkedListNode = reinterpret_cast<char**>(*ptrToNextLinkedListNode);
#define FIVE ONE ONE ONE ONE ONE
#define TEN FIVE FIVE
#define FIFTY TEN TEN TEN TEN TEN
#define HUNDRED FIFTY FIFTY

// Stride is the number of elements to skip when traversing the array.
// It should ideally be >= the cache line to avoid side-effects from pre-fetching.
const int kStrideBytes = 64;

class CacheLatencyTest : public benchmark::Fixture {
    // Fixture for CPU Cache and RAM latency test. Adapted from lmbench's lat_mem_rd test.
public:
    // Array of pointers used as a linked list.
    std::unique_ptr<char*[]> data;

    void SetUp(benchmark::State& state) override {
        if (state.thread_index == 0) {
            fassert(data.get() == nullptr, "'data' is not null");

            /*
             * Create a circular list of pointers using a simple striding
             * algorithm.
             */
            const int arrLength = state.range(0);
            int counter = 0;

            data = std::make_unique<char*[]>(arrLength);

            char** arr = data.get();

            /*
             * This access pattern corresponds to many array/matrix algorithms.
             * It should be easily and correctly predicted by any decent hardware
             * prefetch algorithm.
             */
            for (counter = 0; counter < arrLength - kStrideBytes; counter += kStrideBytes) {
                arr[counter] = reinterpret_cast<char*>(&arr[counter + kStrideBytes]);
            }
            arr[counter] = reinterpret_cast<char*>(&arr[0]);
        }
    }

    void TearDown(benchmark::State& state) override {
        if (state.thread_index == 0) {
            fassert(data.get() != nullptr, "'data' is null");
            data.reset();
        }
    }
};


BENCHMARK_DEFINE_F(CacheLatencyTest, BM_CacheLatency)(benchmark::State& state) {
    size_t arrLength = state.range(0);
    size_t counter = arrLength / (kStrideBytes * 100) + 1;

    for (auto keepRunning : state) {
        char** dummyResult = nullptr;  // Dummy result to prevent the loop from being optimized out.
        char** ptrToNextLinkedListNode = reinterpret_cast<char**>(data.get()[0]);

        for (size_t i = 0; i < counter; ++i) {
            HUNDRED;
        }
        benchmark::DoNotOptimize(dummyResult = ptrToNextLinkedListNode);
    }

    // Record the number of times we accessed the cache so Benchmark can compute the average latency
    // of each access. This allows comparing access latency across caches of different sizes.
    state.SetItemsProcessed(state.iterations() * counter * 100);
}

BENCHMARK_REGISTER_F(CacheLatencyTest, BM_CacheLatency)
    ->RangeMultiplier(2 * 1024)
    // Loop over arrays of different sizes to test the L2, L3, and RAM latency.
    ->Range(256 * 1024, 4096 * 1024)
    ->ThreadRange(1, ProcessInfo::getNumAvailableCores());

}  // namespace
}  // namespace mongo