simdjson/benchmark/benchfeatures.cpp

481 lines
20 KiB
C++

#include "event_counter.h"
#include <cassert>
#include <cctype>
#ifndef _MSC_VER
#include <dirent.h>
#include <unistd.h>
#endif
#include <cinttypes>
#include <initializer_list>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <algorithm>
#include <chrono>
#include <cstring>
#include <fstream>
#include <iomanip>
#include <iostream>
#include <map>
#include <set>
#include <sstream>
#include <string>
#include <vector>
#include "linux-perf-events.h"
#ifdef __linux__
#include <libgen.h>
#endif
#include "simdjson.h"
#include <functional>
#include "benchmarker.h"
using namespace simdjson;
using std::cerr;
using std::cout;
using std::endl;
using std::string;
using std::to_string;
using std::vector;
using std::ostream;
using std::ofstream;
using std::exception;
// Stash the exe_name in main() for functions to use
char* exe_name;
void print_usage(ostream& out) {
out << "Usage: " << exe_name << " [-v] [-n #] [-s STAGE] [-a ARCH]" << endl;
out << endl;
out << "Runs the parser against jsonexamples/generated json files in a loop, measuring speed and other statistics." << endl;
out << endl;
out << "Options:" << endl;
out << endl;
out << "-n # - Number of iterations per file. Default: 400" << endl;
out << "-i # - Number of times to iterate a single file before moving to the next. Default: 20" << endl;
out << "-v - Verbose output." << endl;
out << "-s STAGE - Stop after the given stage." << endl;
out << " -s stage1 - Stop after find_structural_bits." << endl;
out << " -s all - Run all stages." << endl;
out << "-a ARCH - Use the parser with the designated architecture (HASWELL, WESTMERE" << endl;
out << " or ARM64). By default, detects best supported architecture." << endl;
}
void exit_usage(string message) {
cerr << message << endl;
cerr << endl;
print_usage(cerr);
exit(EXIT_FAILURE);
}
struct option_struct {
architecture arch = architecture::UNSUPPORTED;
bool stage1_only = false;
int32_t iterations = 400;
int32_t iteration_step = 50;
bool verbose = false;
option_struct(int argc, char **argv) {
#ifndef _MSC_VER
int c;
while ((c = getopt(argc, argv, "vtn:i:a:s:")) != -1) {
switch (c) {
case 'n':
iterations = atoi(optarg);
break;
case 'i':
iteration_step = atoi(optarg);
break;
case 'v':
verbose = true;
break;
case 'a':
arch = parse_architecture(optarg);
if (arch == architecture::UNSUPPORTED) {
exit_usage(string("Unsupported option value -a ") + optarg + ": expected -a HASWELL, WESTMERE or ARM64");
}
break;
case 's':
if (!strcmp(optarg, "stage1")) {
stage1_only = true;
} else if (!strcmp(optarg, "all")) {
stage1_only = false;
} else {
exit_usage(string("Unsupported option value -s ") + optarg + ": expected -s stage1 or all");
}
break;
default:
exit_error("Unexpected argument " + c);
}
}
#else
int optind = 1;
#endif
// If architecture is not specified, pick the best supported architecture by default
if (arch == architecture::UNSUPPORTED) {
arch = find_best_supported_architecture();
}
document::parser::use_implementation(arch);
}
template<typename F>
void each_stage(const F& f) const {
f(BenchmarkStage::STAGE1);
if (!this->stage1_only) {
f(BenchmarkStage::STAGE2);
f(BenchmarkStage::ALL);
}
}
};
struct feature_benchmarker {
benchmarker utf8;
benchmarker utf8_miss;
benchmarker escape;
benchmarker escape_miss;
benchmarker empty;
benchmarker empty_miss;
benchmarker struct7;
benchmarker struct7_miss;
benchmarker struct7_full;
benchmarker struct15;
benchmarker struct15_miss;
benchmarker struct23;
benchmarker struct23_miss;
feature_benchmarker(const simdjson::implementation &parser, event_collector& collector) :
utf8 ("jsonexamples/generated/utf-8.json", parser, collector),
utf8_miss ("jsonexamples/generated/utf-8-miss.json", parser, collector),
escape ("jsonexamples/generated/escape.json", parser, collector),
escape_miss ("jsonexamples/generated/escape-miss.json", parser, collector),
empty ("jsonexamples/generated/0-structurals.json", parser, collector),
empty_miss ("jsonexamples/generated/0-structurals-miss.json", parser, collector),
struct7 ("jsonexamples/generated/7-structurals.json", parser, collector),
struct7_miss ("jsonexamples/generated/7-structurals-miss.json", parser, collector),
struct7_full ("jsonexamples/generated/7-structurals-full.json", parser, collector),
struct15 ("jsonexamples/generated/15-structurals.json", parser, collector),
struct15_miss("jsonexamples/generated/15-structurals-miss.json", parser, collector),
struct23 ("jsonexamples/generated/23-structurals.json", parser, collector),
struct23_miss("jsonexamples/generated/23-structurals-miss.json", parser, collector)
{
}
really_inline void run_iterations(size_t iterations, bool stage1_only=false) {
struct7.run_iterations(iterations, stage1_only);
struct7_miss.run_iterations(iterations, stage1_only);
struct7_full.run_iterations(iterations, stage1_only);
utf8.run_iterations(iterations, stage1_only);
utf8_miss.run_iterations(iterations, stage1_only);
escape.run_iterations(iterations, stage1_only);
escape_miss.run_iterations(iterations, stage1_only);
empty.run_iterations(iterations, stage1_only);
empty_miss.run_iterations(iterations, stage1_only);
struct15.run_iterations(iterations, stage1_only);
struct15_miss.run_iterations(iterations, stage1_only);
struct23.run_iterations(iterations, stage1_only);
struct23_miss.run_iterations(iterations, stage1_only);
}
double cost_per_block(BenchmarkStage stage, const benchmarker& feature, size_t feature_blocks, const benchmarker& base) const {
return (feature[stage].best.elapsed_ns() - base[stage].best.elapsed_ns()) / feature_blocks;
}
// Whether we're recording cache miss and branch miss events
bool has_events() const {
return empty.collector.has_events();
}
// Base cost of any block (including empty ones)
double base_cost(BenchmarkStage stage) const {
return (empty[stage].best.elapsed_ns() / empty.stats->blocks);
}
// Extra cost of a 1-7 structural block over an empty block
double struct1_7_cost(BenchmarkStage stage) const {
return cost_per_block(stage, struct7, struct7.stats->blocks_with_1_structural, empty);
}
// Extra cost of an 1-7-structural miss
double struct1_7_miss_cost(BenchmarkStage stage) const {
return cost_per_block(stage, struct7_miss, struct7_miss.stats->blocks_with_1_structural, struct7);
}
// Rate of 1-7-structural misses per 8-structural flip
double struct1_7_miss_rate(BenchmarkStage stage) const {
if (!has_events()) { return 1; }
return double(struct7_miss[stage].best.branch_misses() - struct7[stage].best.branch_misses()) / struct7_miss.stats->blocks_with_1_structural_flipped;
}
// Extra cost of an 8-15 structural block over a 1-7 structural block
double struct8_15_cost(BenchmarkStage stage) const {
return cost_per_block(stage, struct15, struct15.stats->blocks_with_8_structurals, struct7);
}
// Extra cost of an 8-15-structural miss over a 1-7 miss
double struct8_15_miss_cost(BenchmarkStage stage) const {
return cost_per_block(stage, struct15_miss, struct15_miss.stats->blocks_with_8_structurals_flipped, struct15);
}
// Rate of 8-15-structural misses per 8-structural flip
double struct8_15_miss_rate(BenchmarkStage stage) const {
if (!has_events()) { return 1; }
return double(struct15_miss[stage].best.branch_misses() - struct15[stage].best.branch_misses()) / struct15_miss.stats->blocks_with_8_structurals_flipped;
}
// Extra cost of a 16+-structural block over an 8-15 structural block (actual varies based on # of structurals!)
double struct16_cost(BenchmarkStage stage) const {
return cost_per_block(stage, struct23, struct23.stats->blocks_with_16_structurals, struct15);
}
// Extra cost of a 16-structural miss over an 8-15 miss
double struct16_miss_cost(BenchmarkStage stage) const {
return cost_per_block(stage, struct23_miss, struct23_miss.stats->blocks_with_16_structurals_flipped, struct23);
}
// Rate of 16-structural misses per 16-structural flip
double struct16_miss_rate(BenchmarkStage stage) const {
if (!has_events()) { return 1; }
return double(struct23_miss[stage].best.branch_misses() - struct23[stage].best.branch_misses()) / struct23_miss.stats->blocks_with_16_structurals_flipped;
}
// Extra cost of having UTF-8 in a block
double utf8_cost(BenchmarkStage stage) const {
return cost_per_block(stage, utf8, utf8.stats->blocks_with_utf8, struct7_full);
}
// Extra cost of a UTF-8 miss
double utf8_miss_cost(BenchmarkStage stage) const {
return cost_per_block(stage, utf8_miss, utf8_miss.stats->blocks_with_utf8_flipped, utf8);
}
// Rate of UTF-8 misses per UTF-8 flip
double utf8_miss_rate(BenchmarkStage stage) const {
if (!has_events()) { return 1; }
return double(utf8_miss[stage].best.branch_misses() - utf8[stage].best.branch_misses()) / utf8_miss.stats->blocks_with_utf8_flipped;
}
// Extra cost of having escapes in a block
double escape_cost(BenchmarkStage stage) const {
return cost_per_block(stage, escape, escape.stats->blocks_with_escapes, struct7_full);
}
// Extra cost of an escape miss
double escape_miss_cost(BenchmarkStage stage) const {
return cost_per_block(stage, escape_miss, escape_miss.stats->blocks_with_escapes_flipped, escape);
}
// Rate of escape misses per escape flip
double escape_miss_rate(BenchmarkStage stage) const {
if (!has_events()) { return 1; }
return double(escape_miss[stage].best.branch_misses() - escape[stage].best.branch_misses()) / escape_miss.stats->blocks_with_escapes_flipped;
}
double calc_expected_feature_cost(BenchmarkStage stage, const benchmarker& file) const {
// Expected base ns/block (empty)
json_stats& stats = *file.stats;
double expected = base_cost(stage) * stats.blocks;
expected += struct1_7_cost(stage) * stats.blocks_with_1_structural;
expected += utf8_cost(stage) * stats.blocks_with_utf8;
expected += escape_cost(stage) * stats.blocks_with_escapes;
expected += struct8_15_cost(stage) * stats.blocks_with_8_structurals;
expected += struct16_cost(stage) * stats.blocks_with_16_structurals;
return expected / stats.blocks;
}
double calc_expected_miss_cost(BenchmarkStage stage, const benchmarker& file) const {
// Expected base ns/block (empty)
json_stats& stats = *file.stats;
double expected = struct1_7_miss_cost(stage) * stats.blocks_with_1_structural_flipped * struct1_7_miss_rate(stage);
expected += utf8_miss_cost(stage) * stats.blocks_with_utf8_flipped * utf8_miss_rate(stage);
expected += escape_miss_cost(stage) * stats.blocks_with_escapes_flipped * escape_miss_rate(stage);
expected += struct8_15_miss_cost(stage) * stats.blocks_with_8_structurals_flipped * struct8_15_miss_rate(stage);
expected += struct16_miss_cost(stage) * stats.blocks_with_16_structurals_flipped * struct16_miss_rate(stage);
return expected / stats.blocks;
}
double calc_expected_misses(BenchmarkStage stage, const benchmarker& file) const {
json_stats& stats = *file.stats;
double expected = stats.blocks_with_1_structural_flipped * struct1_7_miss_rate(stage);
expected += stats.blocks_with_utf8_flipped * utf8_miss_rate(stage);
expected += stats.blocks_with_escapes_flipped * escape_miss_rate(stage);
expected += stats.blocks_with_8_structurals_flipped * struct8_15_miss_rate(stage);
expected += stats.blocks_with_16_structurals_flipped * struct16_miss_rate(stage);
return expected;
}
double calc_expected(BenchmarkStage stage, const benchmarker& file) const {
return calc_expected_feature_cost(stage, file) + calc_expected_miss_cost(stage, file);
}
void print(const option_struct& options) const {
printf("\n");
printf("Features in ns/block (64 bytes):\n");
printf("\n");
printf("| %-8s ", "Stage");
printf("| %8s ", "Base");
printf("| %8s ", "7 Struct");
printf("| %8s ", "UTF-8");
printf("| %8s ", "Escape");
printf("| %8s ", "15 Str.");
printf("| %8s ", "16+ Str.");
printf("| %15s ", "7 Struct Miss");
printf("| %15s ", "UTF-8 Miss");
printf("| %15s ", "Escape Miss");
printf("| %15s ", "15 Str. Miss");
printf("| %15s ", "16+ Str. Miss");
printf("|\n");
printf("|%.10s", "---------------------------------------");
printf("|%.10s", "---------------------------------------");
printf("|%.10s", "---------------------------------------");
printf("|%.10s", "---------------------------------------");
printf("|%.10s", "---------------------------------------");
printf("|%.10s", "---------------------------------------");
printf("|%.10s", "---------------------------------------");
printf("|%.17s", "---------------------------------------");
printf("|%.17s", "---------------------------------------");
printf("|%.17s", "---------------------------------------");
printf("|%.17s", "---------------------------------------");
printf("|%.17s", "---------------------------------------");
printf("|\n");
options.each_stage([&](auto stage) {
printf("| %-8s ", benchmark_stage_name(stage));
printf("| %8.3g ", base_cost(stage));
printf("| %8.3g ", struct1_7_cost(stage));
printf("| %8.3g ", utf8_cost(stage));
printf("| %8.3g ", escape_cost(stage));
printf("| %8.3g ", struct8_15_cost(stage));
printf("| %8.3g ", struct16_cost(stage));
if (has_events()) {
printf("| %8.3g (%3d%%) ", struct1_7_miss_cost(stage), int(struct1_7_miss_rate(stage)*100));
printf("| %8.3g (%3d%%) ", utf8_miss_cost(stage), int(utf8_miss_rate(stage)*100));
printf("| %8.3g (%3d%%) ", escape_miss_cost(stage), int(escape_miss_rate(stage)*100));
printf("| %8.3g (%3d%%) ", struct8_15_miss_cost(stage), int(struct8_15_miss_rate(stage)*100));
printf("| %8.3g (%3d%%) ", struct16_miss_cost(stage), int(struct16_miss_rate(stage)*100));
} else {
printf("| %8.3g ", struct1_7_miss_cost(stage));
printf("| %8.3g ", utf8_miss_cost(stage));
printf("| %8.3g ", escape_miss_cost(stage));
printf("| %8.3g ", struct8_15_miss_cost(stage));
printf("| %8.3g ", struct16_miss_cost(stage));
}
printf("|\n");
});
}
};
void print_file_effectiveness(BenchmarkStage stage, const char* filename, const benchmarker& results, const feature_benchmarker& features) {
double actual = results[stage].best.elapsed_ns() / results.stats->blocks;
double calc = features.calc_expected(stage, results);
uint64_t actual_misses = results[stage].best.branch_misses();
uint64_t calc_misses = uint64_t(features.calc_expected_misses(stage, results));
double calc_miss_cost = features.calc_expected_miss_cost(stage, results);
printf(" | %-8s ", benchmark_stage_name(stage));
printf("| %-15s ", filename);
printf("| %8.3g ", features.calc_expected_feature_cost(stage, results));
printf("| %8.3g ", calc_miss_cost);
printf("| %8.3g ", calc);
printf("| %8.3g ", actual);
printf("| %+8.3g ", actual - calc);
printf("| %13lu ", calc_misses);
if (features.has_events()) {
printf("| %13lu ", actual_misses);
printf("| %+13ld ", int64_t(actual_misses - calc_misses));
double miss_adjustment = calc_miss_cost * (double(int64_t(actual_misses - calc_misses)) / calc_misses);
printf("| %8.3g ", calc_miss_cost + miss_adjustment);
printf("| %+8.3g ", actual - (calc + miss_adjustment));
}
printf("|\n");
}
int main(int argc, char *argv[]) {
// Read options
exe_name = argv[0];
option_struct options(argc, argv);
if (options.verbose) {
verbose_stream = &cout;
}
// Initialize the event collector. We put this early so if it prints an error message, it's the
// first thing printed.
event_collector collector;
// Set up benchmarkers by reading all files
feature_benchmarker features(collector);
benchmarker gsoc_2018("jsonexamples/gsoc-2018.json", collector);
benchmarker twitter("jsonexamples/twitter.json", collector);
benchmarker random("jsonexamples/random.json", collector);
// Run the benchmarks
progress_bar progress(options.iterations, 100);
// Put the if (options.stage1_only) *outside* the loop so that run_iterations will be optimized
if (options.stage1_only) {
for (int iteration = 0; iteration < options.iterations; iteration += options.iteration_step) {
if (!options.verbose) { progress.print(iteration); }
features.run_iterations(options.iteration_step, true);
gsoc_2018.run_iterations(options.iteration_step, true);
twitter.run_iterations(options.iteration_step, true);
random.run_iterations(options.iteration_step, true);
}
} else {
for (int iteration = 0; iteration < options.iterations; iteration += options.iteration_step) {
if (!options.verbose) { progress.print(iteration); }
features.run_iterations(options.iteration_step, false);
gsoc_2018.run_iterations(options.iteration_step, false);
twitter.run_iterations(options.iteration_step, false);
random.run_iterations(options.iteration_step, false);
}
}
if (!options.verbose) { progress.erase(); }
features.print(options);
// Gauge effectiveness
if (options.verbose) {
printf("\n");
printf(" Effectiveness Check: Estimated vs. Actual ns/block for real files:\n");
printf("\n");
printf(" | %8s ", "Stage");
printf("| %-15s ", "File");
printf("| %11s ", "Est. (Base)");
printf("| %11s ", "Est. (Miss)");
printf("| %8s ", "Est.");
printf("| %8s ", "Actual");
printf("| %8s ", "Diff");
printf("| %13s ", "Est. Misses");
if (features.has_events()) {
printf("| %13s ", "Actual Misses");
printf("| %13s ", "Diff (Misses)");
printf("| %13s ", "Adjusted Miss");
printf("| %13s ", "Adjusted Diff");
}
printf("|\n");
printf(" |%.10s", "---------------------------------------");
printf("|%.17s", "---------------------------------------");
printf("|%.13s", "---------------------------------------");
printf("|%.13s", "---------------------------------------");
printf("|%.10s", "---------------------------------------");
printf("|%.10s", "---------------------------------------");
printf("|%.10s", "---------------------------------------");
printf("|%.15s", "---------------------------------------");
if (features.has_events()) {
printf("|%.15s", "---------------------------------------");
printf("|%.15s", "---------------------------------------");
printf("|%.15s", "---------------------------------------");
printf("|%.15s", "---------------------------------------");
}
printf("|\n");
options.each_stage([&](auto stage) {
print_file_effectiveness(stage, "gsoc-2018.json", gsoc_2018, features);
print_file_effectiveness(stage, "twitter.json", twitter, features);
print_file_effectiveness(stage, "random.json", random, features);
});
}
return EXIT_SUCCESS;
}