2 * Copyright 2017 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 // @author Andrei Alexandrescu (andrei.alexandrescu@fb.com)
19 #include <folly/Benchmark.h>
20 #include <folly/Foreach.h>
21 #include <folly/json.h>
22 #include <folly/String.h>
25 #include <boost/regex.hpp>
35 DEFINE_bool(benchmark, false, "Run benchmarks.");
36 DEFINE_bool(json, false, "Output in JSON format.");
41 "Only benchmarks whose names match this regex will be run.");
46 "Minimum # of microseconds we'll accept for each benchmark.");
51 "Minimum # of iterations we'll try for each benchmark.");
56 "Maximum # of iterations we'll try for each benchmark.");
61 "Maximum # of seconds we'll spend on each benchmark.");
65 std::chrono::high_resolution_clock::duration BenchmarkSuspender::timeSpent;
67 typedef function<detail::TimeIterPair(unsigned int)> BenchmarkFun;
70 vector<tuple<string, string, BenchmarkFun>>& benchmarks() {
71 static vector<tuple<string, string, BenchmarkFun>> _benchmarks;
75 #define FB_FOLLY_GLOBAL_BENCHMARK_BASELINE fbFollyGlobalBenchmarkBaseline
76 #define FB_STRINGIZE_X2(x) FB_STRINGIZE(x)
78 // Add the global baseline
79 BENCHMARK(FB_FOLLY_GLOBAL_BENCHMARK_BASELINE) {
87 size_t getGlobalBenchmarkBaselineIndex() {
88 const char *global = FB_STRINGIZE_X2(FB_FOLLY_GLOBAL_BENCHMARK_BASELINE);
89 auto it = std::find_if(
92 [global](const tuple<string, string, BenchmarkFun> &v) {
93 return get<1>(v) == global;
96 CHECK(it != benchmarks().end());
97 return size_t(std::distance(benchmarks().begin(), it));
100 #undef FB_STRINGIZE_X2
101 #undef FB_FOLLY_GLOBAL_BENCHMARK_BASELINE
103 void detail::addBenchmarkImpl(const char* file, const char* name,
105 benchmarks().emplace_back(file, name, std::move(fun));
109 * Given a bunch of benchmark samples, estimate the actual run time.
111 static double estimateTime(double * begin, double * end) {
114 // Current state of the art: get the minimum. After some
115 // experimentation, it seems taking the minimum is the best.
116 return *min_element(begin, end);
119 static double runBenchmarkGetNSPerIteration(const BenchmarkFun& fun,
120 const double globalBaseline) {
121 using std::chrono::duration_cast;
122 using std::chrono::high_resolution_clock;
123 using std::chrono::microseconds;
124 using std::chrono::nanoseconds;
125 using std::chrono::seconds;
127 // They key here is accuracy; too low numbers means the accuracy was
128 // coarse. We up the ante until we get to at least minNanoseconds
131 std::is_same<high_resolution_clock::duration, nanoseconds>::value,
132 "High resolution clock must be nanosecond resolution.");
133 // We choose a minimum minimum (sic) of 100,000 nanoseconds, but if
134 // the clock resolution is worse than that, it will be larger. In
135 // essence we're aiming at making the quantization noise 0.01%.
136 static const auto minNanoseconds = std::max<nanoseconds>(
137 nanoseconds(100000), microseconds(FLAGS_bm_min_usec));
139 // We do measurements in several epochs and take the minimum, to
140 // account for jitter.
141 static const unsigned int epochs = 1000;
142 // We establish a total time budget as we don't want a measurement
143 // to take too long. This will curtail the number of actual epochs.
144 const auto timeBudget = seconds(FLAGS_bm_max_secs);
145 auto global = high_resolution_clock::now();
147 double epochResults[epochs] = { 0 };
148 size_t actualEpochs = 0;
150 for (; actualEpochs < epochs; ++actualEpochs) {
151 const auto maxIters = uint32_t(FLAGS_bm_max_iters);
152 for (auto n = uint32_t(FLAGS_bm_min_iters); n < maxIters; n *= 2) {
153 auto const nsecsAndIter = fun(static_cast<unsigned int>(n));
154 if (nsecsAndIter.first < minNanoseconds) {
157 // We got an accurate enough timing, done. But only save if
158 // smaller than the current result.
159 auto nsecs = duration_cast<nanoseconds>(nsecsAndIter.first).count();
160 epochResults[actualEpochs] =
161 max(0.0, double(nsecs) / nsecsAndIter.second - globalBaseline);
162 // Done with the current epoch, we got a meaningful timing.
165 auto now = high_resolution_clock::now();
166 if (now - global >= timeBudget) {
167 // No more time budget available.
173 // If the benchmark was basically drowned in baseline noise, it's
174 // possible it became negative.
175 return max(0.0, estimateTime(epochResults, epochResults + actualEpochs));
183 static const ScaleInfo kTimeSuffixes[] {
184 { 365.25 * 24 * 3600, "years" },
185 { 24 * 3600, "days" },
197 static const ScaleInfo kMetricSuffixes[] {
198 { 1E24, "Y" }, // yotta
199 { 1E21, "Z" }, // zetta
200 { 1E18, "X" }, // "exa" written with suffix 'X' so as to not create
201 // confusion with scientific notation
202 { 1E15, "P" }, // peta
203 { 1E12, "T" }, // terra
204 { 1E9, "G" }, // giga
205 { 1E6, "M" }, // mega
206 { 1E3, "K" }, // kilo
208 { 1E-3, "m" }, // milli
209 { 1E-6, "u" }, // micro
210 { 1E-9, "n" }, // nano
211 { 1E-12, "p" }, // pico
212 { 1E-15, "f" }, // femto
213 { 1E-18, "a" }, // atto
214 { 1E-21, "z" }, // zepto
215 { 1E-24, "y" }, // yocto
219 static string humanReadable(double n, unsigned int decimals,
220 const ScaleInfo* scales) {
221 if (std::isinf(n) || std::isnan(n)) {
222 return folly::to<string>(n);
225 const double absValue = fabs(n);
226 const ScaleInfo* scale = scales;
227 while (absValue < scale[0].boundary && scale[1].suffix != nullptr) {
231 const double scaledValue = n / scale->boundary;
232 return stringPrintf("%.*f%s", decimals, scaledValue, scale->suffix);
235 static string readableTime(double n, unsigned int decimals) {
236 return humanReadable(n, decimals, kTimeSuffixes);
239 static string metricReadable(double n, unsigned int decimals) {
240 return humanReadable(n, decimals, kMetricSuffixes);
243 static void printBenchmarkResultsAsTable(
244 const vector<tuple<string, string, double> >& data) {
246 static const unsigned int columns = 76;
248 // Compute the longest benchmark name
249 size_t longestName = 0;
250 FOR_EACH_RANGE (i, 1, benchmarks().size()) {
251 longestName = max(longestName, get<1>(benchmarks()[i]).size());
254 // Print a horizontal rule
255 auto separator = [&](char pad) {
256 puts(string(columns, pad).c_str());
259 // Print header for a file
260 auto header = [&](const string& file) {
262 printf("%-*srelative time/iter iters/s\n",
263 columns - 28, file.c_str());
267 double baselineNsPerIter = numeric_limits<double>::max();
270 for (auto& datum : data) {
271 auto file = get<0>(datum);
272 if (file != lastFile) {
278 string s = get<1>(datum);
283 bool useBaseline /* = void */;
288 baselineNsPerIter = get<2>(datum);
291 s.resize(columns - 29, ' ');
292 auto nsPerIter = get<2>(datum);
293 auto secPerIter = nsPerIter / 1E9;
294 auto itersPerSec = (secPerIter == 0)
295 ? std::numeric_limits<double>::infinity()
298 // Print without baseline
299 printf("%*s %9s %7s\n",
300 static_cast<int>(s.size()), s.c_str(),
301 readableTime(secPerIter, 2).c_str(),
302 metricReadable(itersPerSec, 2).c_str());
304 // Print with baseline
305 auto rel = baselineNsPerIter / nsPerIter * 100.0;
306 printf("%*s %7.2f%% %9s %7s\n",
307 static_cast<int>(s.size()), s.c_str(),
309 readableTime(secPerIter, 2).c_str(),
310 metricReadable(itersPerSec, 2).c_str());
316 static void printBenchmarkResultsAsJson(
317 const vector<tuple<string, string, double> >& data) {
318 dynamic d = dynamic::object;
319 for (auto& datum: data) {
320 d[std::get<1>(datum)] = std::get<2>(datum) * 1000.;
323 printf("%s\n", toPrettyJson(d).c_str());
326 static void printBenchmarkResults(
327 const vector<tuple<string, string, double> >& data) {
330 printBenchmarkResultsAsJson(data);
332 printBenchmarkResultsAsTable(data);
336 void runBenchmarks() {
337 CHECK(!benchmarks().empty());
339 vector<tuple<string, string, double>> results;
340 results.reserve(benchmarks().size() - 1);
342 std::unique_ptr<boost::regex> bmRegex;
343 if (!FLAGS_bm_regex.empty()) {
344 bmRegex.reset(new boost::regex(FLAGS_bm_regex));
347 // PLEASE KEEP QUIET. MEASUREMENTS IN PROGRESS.
349 size_t baselineIndex = getGlobalBenchmarkBaselineIndex();
351 auto const globalBaseline =
352 runBenchmarkGetNSPerIteration(get<2>(benchmarks()[baselineIndex]), 0);
353 FOR_EACH_RANGE (i, 0, benchmarks().size()) {
354 if (i == baselineIndex) {
357 double elapsed = 0.0;
358 if (get<1>(benchmarks()[i]) != "-") { // skip separators
359 if (bmRegex && !boost::regex_search(get<1>(benchmarks()[i]), *bmRegex)) {
362 elapsed = runBenchmarkGetNSPerIteration(get<2>(benchmarks()[i]),
365 results.emplace_back(get<0>(benchmarks()[i]),
366 get<1>(benchmarks()[i]), elapsed);
369 // PLEASE MAKE NOISE. MEASUREMENTS DONE.
371 printBenchmarkResults(results);