2 * Copyright 2014 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 "Benchmark.h"
25 #include <boost/regex.hpp>
34 DEFINE_bool(benchmark, false, "Run benchmarks.");
35 DEFINE_bool(json, false, "Output in JSON format.");
37 DEFINE_string(bm_regex, "",
38 "Only benchmarks whose names match this regex will be run.");
40 DEFINE_int64(bm_min_usec, 100,
41 "Minimum # of microseconds we'll accept for each benchmark.");
43 DEFINE_int64(bm_min_iters, 1,
44 "Minimum # of iterations we'll try for each benchmark.");
46 DEFINE_int32(bm_max_secs, 1,
47 "Maximum # of seconds we'll spend on each benchmark.");
52 BenchmarkSuspender::NanosecondsSpent BenchmarkSuspender::nsSpent;
54 typedef function<uint64_t(unsigned int)> BenchmarkFun;
55 static vector<tuple<const char*, const char*, BenchmarkFun>> benchmarks;
57 // Add the global baseline
58 BENCHMARK(globalBenchmarkBaseline) {
66 void detail::addBenchmarkImpl(const char* file, const char* name,
68 benchmarks.emplace_back(file, name, std::move(fun));
72 * Given a point, gives density at that point as a number 0.0 < x <=
73 * 1.0. The result is 1.0 if all samples are equal to where, and
74 * decreases near 0 if all points are far away from it. The density is
75 * computed with the help of a radial basis function.
77 static double density(const double * begin, const double *const end,
78 const double where, const double bandwidth) {
80 assert(bandwidth > 0.0);
82 FOR_EACH_RANGE (i, begin, end) {
83 auto d = (*i - where) / bandwidth;
86 return sum / (end - begin);
90 * Computes mean and variance for a bunch of data points. Note that
91 * mean is currently not being used.
93 static pair<double, double>
94 meanVariance(const double * begin, const double *const end) {
96 double sum = 0.0, sum2 = 0.0;
97 FOR_EACH_RANGE (i, begin, end) {
101 auto const n = end - begin;
102 return make_pair(sum / n, sqrt((sum2 - sum * sum / n) / n));
106 * Computes the mode of a sample set through brute force. Assumes
109 static double mode(const double * begin, const double *const end) {
111 // Lower bound and upper bound for result and their respective
117 // Get the variance so we pass it down to density()
118 auto const sigma = meanVariance(begin, end).second;
120 // No variance means constant signal
124 FOR_EACH_RANGE (i, begin, end) {
125 assert(i == begin || *i >= i[-1]);
126 auto candidate = density(begin, end, *i, sigma * sqrt(2.0));
127 if (candidate > bestDensity) {
129 bestDensity = candidate;
132 // Density is decreasing... we could break here if we definitely
133 // knew this is unimodal.
141 * Given a bunch of benchmark samples, estimate the actual run time.
143 static double estimateTime(double * begin, double * end) {
146 // Current state of the art: get the minimum. After some
147 // experimentation, it seems taking the minimum is the best.
149 return *min_element(begin, end);
151 // What follows after estimates the time as the mode of the
154 // Select the awesomest (i.e. most frequent) result. We do this by
155 // sorting and then computing the longest run length.
158 // Eliminate outliers. A time much larger than the minimum time is
159 // considered an outlier.
160 while (end[-1] > 2.0 * *begin) {
170 /* Code used just for comparison purposes */ {
171 unsigned bestFrequency = 0;
172 unsigned candidateFrequency = 1;
173 double candidateValue = *begin;
174 for (auto current = begin + 1; ; ++current) {
175 if (current == end || *current != candidateValue) {
176 // Done with the current run, see if it was best
177 if (candidateFrequency > bestFrequency) {
178 bestFrequency = candidateFrequency;
179 result = candidateValue;
181 if (current == end) {
185 candidateValue = *current;
186 candidateFrequency = 1;
188 // Cool, inside a run, increase the frequency
189 ++candidateFrequency;
194 result = mode(begin, end);
199 static double runBenchmarkGetNSPerIteration(const BenchmarkFun& fun,
200 const double globalBaseline) {
201 // They key here is accuracy; too low numbers means the accuracy was
202 // coarse. We up the ante until we get to at least minNanoseconds
204 static uint64_t resolutionInNs = 0, coarseResolutionInNs = 0;
205 if (!resolutionInNs) {
207 CHECK_EQ(0, clock_getres(detail::DEFAULT_CLOCK_ID, &ts));
208 CHECK_EQ(0, ts.tv_sec) << "Clock sucks.";
209 CHECK_LT(0, ts.tv_nsec) << "Clock too fast for its own good.";
210 CHECK_EQ(1, ts.tv_nsec) << "Clock too coarse, upgrade your kernel.";
211 resolutionInNs = ts.tv_nsec;
213 // We choose a minimum minimum (sic) of 100,000 nanoseconds, but if
214 // the clock resolution is worse than that, it will be larger. In
215 // essence we're aiming at making the quantization noise 0.01%.
216 static const auto minNanoseconds =
217 max<uint64_t>(FLAGS_bm_min_usec * 1000UL,
218 min<uint64_t>(resolutionInNs * 100000, 1000000000ULL));
220 // We do measurements in several epochs and take the minimum, to
221 // account for jitter.
222 static const unsigned int epochs = 1000;
223 // We establish a total time budget as we don't want a measurement
224 // to take too long. This will curtail the number of actual epochs.
225 const uint64_t timeBudgetInNs = FLAGS_bm_max_secs * 1000000000;
227 CHECK_EQ(0, clock_gettime(CLOCK_REALTIME, &global));
229 double epochResults[epochs] = { 0 };
230 size_t actualEpochs = 0;
232 for (; actualEpochs < epochs; ++actualEpochs) {
233 for (unsigned int n = FLAGS_bm_min_iters; n < (1UL << 30); n *= 2) {
234 auto const nsecs = fun(n);
235 if (nsecs < minNanoseconds) {
238 // We got an accurate enough timing, done. But only save if
239 // smaller than the current result.
240 epochResults[actualEpochs] = max(0.0, double(nsecs) / n - globalBaseline);
241 // Done with the current epoch, we got a meaningful timing.
245 CHECK_EQ(0, clock_gettime(CLOCK_REALTIME, &now));
246 if (detail::timespecDiff(now, global) >= timeBudgetInNs) {
247 // No more time budget available.
253 // If the benchmark was basically drowned in baseline noise, it's
254 // possible it became negative.
255 return max(0.0, estimateTime(epochResults, epochResults + actualEpochs));
263 static const ScaleInfo kTimeSuffixes[] {
264 { 365.25 * 24 * 3600, "years" },
265 { 24 * 3600, "days" },
277 static const ScaleInfo kMetricSuffixes[] {
278 { 1E24, "Y" }, // yotta
279 { 1E21, "Z" }, // zetta
280 { 1E18, "X" }, // "exa" written with suffix 'X' so as to not create
281 // confusion with scientific notation
282 { 1E15, "P" }, // peta
283 { 1E12, "T" }, // terra
284 { 1E9, "G" }, // giga
285 { 1E6, "M" }, // mega
286 { 1E3, "K" }, // kilo
288 { 1E-3, "m" }, // milli
289 { 1E-6, "u" }, // micro
290 { 1E-9, "n" }, // nano
291 { 1E-12, "p" }, // pico
292 { 1E-15, "f" }, // femto
293 { 1E-18, "a" }, // atto
294 { 1E-21, "z" }, // zepto
295 { 1E-24, "y" }, // yocto
299 static string humanReadable(double n, unsigned int decimals,
300 const ScaleInfo* scales) {
301 if (std::isinf(n) || std::isnan(n)) {
302 return folly::to<string>(n);
305 const double absValue = fabs(n);
306 const ScaleInfo* scale = scales;
307 while (absValue < scale[0].boundary && scale[1].suffix != nullptr) {
311 const double scaledValue = n / scale->boundary;
312 return stringPrintf("%.*f%s", decimals, scaledValue, scale->suffix);
315 static string readableTime(double n, unsigned int decimals) {
316 return humanReadable(n, decimals, kTimeSuffixes);
319 static string metricReadable(double n, unsigned int decimals) {
320 return humanReadable(n, decimals, kMetricSuffixes);
323 static void printBenchmarkResultsAsTable(
324 const vector<tuple<const char*, const char*, double> >& data) {
326 static const unsigned int columns = 76;
328 // Compute the longest benchmark name
329 size_t longestName = 0;
330 FOR_EACH_RANGE (i, 1, benchmarks.size()) {
331 longestName = max(longestName, strlen(get<1>(benchmarks[i])));
334 // Print a horizontal rule
335 auto separator = [&](char pad) {
336 puts(string(columns, pad).c_str());
339 // Print header for a file
340 auto header = [&](const char* file) {
342 printf("%-*srelative time/iter iters/s\n",
347 double baselineNsPerIter = numeric_limits<double>::max();
348 const char* lastFile = "";
350 for (auto& datum : data) {
351 auto file = get<0>(datum);
352 if (strcmp(file, lastFile)) {
358 string s = get<1>(datum);
363 bool useBaseline /* = void */;
368 baselineNsPerIter = get<2>(datum);
371 s.resize(columns - 29, ' ');
372 auto nsPerIter = get<2>(datum);
373 auto secPerIter = nsPerIter / 1E9;
374 auto itersPerSec = 1 / secPerIter;
376 // Print without baseline
377 printf("%*s %9s %7s\n",
378 static_cast<int>(s.size()), s.c_str(),
379 readableTime(secPerIter, 2).c_str(),
380 metricReadable(itersPerSec, 2).c_str());
382 // Print with baseline
383 auto rel = baselineNsPerIter / nsPerIter * 100.0;
384 printf("%*s %7.2f%% %9s %7s\n",
385 static_cast<int>(s.size()), s.c_str(),
387 readableTime(secPerIter, 2).c_str(),
388 metricReadable(itersPerSec, 2).c_str());
394 static void printBenchmarkResultsAsJson(
395 const vector<tuple<const char*, const char*, double> >& data) {
396 dynamic d = dynamic::object;
397 for (auto& datum: data) {
398 d[std::get<1>(datum)] = std::get<2>(datum) * 1000.;
401 printf("%s\n", toPrettyJson(d).c_str());
404 static void printBenchmarkResults(
405 const vector<tuple<const char*, const char*, double> >& data) {
408 printBenchmarkResultsAsJson(data);
410 printBenchmarkResultsAsTable(data);
414 void runBenchmarks() {
415 CHECK(!benchmarks.empty());
417 vector<tuple<const char*, const char*, double>> results;
418 results.reserve(benchmarks.size() - 1);
420 std::unique_ptr<boost::regex> bmRegex;
421 if (!FLAGS_bm_regex.empty()) {
422 bmRegex.reset(new boost::regex(FLAGS_bm_regex));
425 // PLEASE KEEP QUIET. MEASUREMENTS IN PROGRESS.
427 auto const globalBaseline = runBenchmarkGetNSPerIteration(
428 get<2>(benchmarks.front()), 0);
429 FOR_EACH_RANGE (i, 1, benchmarks.size()) {
430 double elapsed = 0.0;
431 if (strcmp(get<1>(benchmarks[i]), "-") != 0) { // skip separators
432 if (bmRegex && !boost::regex_search(get<1>(benchmarks[i]), *bmRegex)) {
435 elapsed = runBenchmarkGetNSPerIteration(get<2>(benchmarks[i]),
438 results.emplace_back(get<0>(benchmarks[i]),
439 get<1>(benchmarks[i]), elapsed);
442 // PLEASE MAKE NOISE. MEASUREMENTS DONE.
444 printBenchmarkResults(results);