programs/perf_tests/UnfoldTopology.cc

0001
0002 #include "JANA/Utils/JEventLevel.h"
0003 #include <catch.hpp>
0004
0005 #include <JANA/JApplication.h>
0006 #include <JANA/JFactoryGenerator.h>
0007 #include <JANA/CLI/JBenchmarker.h>
0008 #include <JANA/JEventUnfolder.h>
0009 #include <JANA/JEventProcessor.h>
0010 #include <JANA/JEventSource.h>
0011 #include <JANA/Utils/JBenchUtils.h>
0012
0013
0014 namespace jana::perftest::unfold {
0015
0016 struct Data { size_t x; };
0017
0018 struct BSrc : public JEventSource {
0019
0020     Output<Data> data_out {this};
0021     Parameter<int> latency_us {this, "latency_us", 0};
0022
0023     BSrc() {
0024         SetPrefix("bsrc");
0025         SetLevel(JEventLevel::Block);
0026         SetCallbackStyle(CallbackStyle::ExpertMode);
0027         data_out.SetShortName("1");
0028     }
0029     JEventSource::Result Emit(JEvent& block) override {
0030         data_out().push_back(new Data {block.GetEventNumber()*3 });
0031         JBenchUtils::consume_cpu_us(*latency_us);
0032         return Result::Success;
0033     };
0034 };
0035
0036 struct BFac : public JFactory {
0037     Input<Data> data_in {this};
0038     Output<Data> data_out {this};
0039     Parameter<int> latency_us {this, "latency_us", 0};
0040
0041     BFac() {
0042         SetPrefix("bfac");
0043         SetLevel(JEventLevel::Block);
0044         data_in.SetDatabundleName("1");
0045         data_out.SetShortName("2");
0046     }
0047     void Process(const JEvent&) override {
0048         auto x = data_in().at(0)->x + 7;
0049         data_out().push_back(new Data {x});
0050         JBenchUtils::consume_cpu_us(*latency_us);
0051     };
0052 };
0053
0054 struct Unf : public JEventUnfolder {
0055     Input<Data> parent_data_in {this};
0056     Output<Data> child_data_out {this};
0057     Parameter<int> latency_us {this, "latency_us", 0};
0058
0059     Unf() {
0060         SetPrefix("unf");
0061         SetParentLevel(JEventLevel::Block);
0062         SetChildLevel(JEventLevel::PhysicsEvent);
0063         SetCallbackStyle(CallbackStyle::ExpertMode);
0064         parent_data_in.SetDatabundleName("2");
0065         child_data_out.SetShortName("3");
0066     }
0067     JEventUnfolder::Result Unfold(const JEvent&, JEvent&, int child_nr) override {
0068         auto x = parent_data_in().at(0)->x * 10 + child_nr;
0069         child_data_out().push_back(new Data{x});
0070         JBenchUtils::consume_cpu_us(*latency_us);
0071
0072         if (child_nr == 9) {
0073             return Result::NextChildNextParent;
0074         }
0075         return Result::NextChildKeepParent;
0076     };
0077 };
0078
0079 struct PEFac : public JFactory {
0080     Input<Data> data_in {this};
0081     Output<Data> data_out {this};
0082     Parameter<int> latency_us {this, "latency_us", 0};
0083     PEFac() {
0084         SetPrefix("pefac");
0085         SetLevel(JEventLevel::PhysicsEvent);
0086         data_in.SetDatabundleName("3");
0087         data_out.SetShortName("4");
0088     }
0089     void Process(const JEvent&) override {
0090         auto x = data_in().at(0)->x * 10;
0091         data_out().push_back(new Data {x});
0092         JBenchUtils::consume_cpu_us(*latency_us);
0093     };
0094 };
0095
0096 struct PEProc : public JEventProcessor {
0097     Input<Data> data_in {this};
0098     Parameter<int> latency_us {this, "latency_us", 0};
0099     PEProc() {
0100         SetPrefix("peproc");
0101         SetLevel(JEventLevel::PhysicsEvent);
0102         SetCallbackStyle(CallbackStyle::ExpertMode);
0103         data_in.SetDatabundleName("4");
0104     }
0105     void ProcessSequential(const JEvent&) override {
0106         (void) (data_in().at(0)->x);
0107         JBenchUtils::consume_cpu_us(*latency_us);
0108     };
0109 };
0110
0111
0112 TEST_CASE("UnfoldTopology_Mini") {
0113
0114     LOG << "Running UnfoldTopology_Mini";
0115
0116     JApplication app;
0117     app.SetParameterValue("bsrc:latency_us", 0); // Infinity Hz
0118     app.SetParameterValue("bfac:latency_us", 0); // Infinity Hz
0119     app.SetParameterValue("unf:latency_us", 0); // Infinity Hz
0120     app.SetParameterValue("pefac:latency_us", 0); // Infinity Hz
0121     app.SetParameterValue("peproc:latency_us", 0); // Infinity Hz
0122     app.SetParameterValue("benchmark:resultsdir", "docs/perf_tests");
0123     app.SetParameterValue("benchmark:rates_filename", "unfold_mini.dat");
0124     app.SetParameterValue("benchmark:use_log_scale", true);
0125     app.SetParameterValue("benchmark:minthreads", "1");
0126     app.SetParameterValue("benchmark:maxthreads", "32");
0127
0128     app.Add(new BSrc);
0129     app.Add(new Unf);
0130     app.Add(new PEProc);
0131     app.Add(new JFactoryGeneratorT<BFac>);
0132     app.Add(new JFactoryGeneratorT<PEFac>);
0133
0134     JBenchmarker benchmarker(&app);
0135     benchmarker.RunUntilFinished();
0136 }
0137
0138 TEST_CASE("UnfoldTopology_Small") {
0139     LOG << "Running UnfoldTopology_Small";
0140
0141     JApplication app;
0142     app.SetParameterValue("bsrc:latency_us", 0); // Infinity Hz
0143     app.SetParameterValue("bfac:latency_us", 1'000'000 / 500); // 500 Hz
0144     app.SetParameterValue("unf:latency_us", 0); // Infinity Hz
0145     app.SetParameterValue("pefac:latency_us", 1'000'000 / 5000); // 5 kHz
0146     app.SetParameterValue("peproc:latency_us", 0); // Infinity Hz
0147     app.SetParameterValue("benchmark:resultsdir", "docs/perf_tests");
0148     app.SetParameterValue("benchmark:rates_filename", "unfold_small.dat");
0149     app.SetParameterValue("benchmark:use_log_scale", true);
0150     app.SetParameterValue("benchmark:minthreads", "1");
0151     app.SetParameterValue("benchmark:maxthreads", "32");
0152
0153     app.Add(new BSrc);
0154     app.Add(new Unf);
0155     app.Add(new PEProc);
0156     app.Add(new JFactoryGeneratorT<BFac>);
0157     app.Add(new JFactoryGeneratorT<PEFac>);
0158
0159     JBenchmarker benchmarker(&app);
0160     benchmarker.RunUntilFinished();
0161 }
0162
0163 TEST_CASE("UnfoldTopology_Small_Saturation") {
0164     LOG << "Running UnfoldTopology_Small_Saturation";
0165
0166     JApplication app;
0167     app.SetParameterValue("bsrc:latency_us", 1'000'000 / 2'000); // 2 kHz
0168     app.SetParameterValue("bfac:latency_us", 1'000'000 / 500); // 500 Hz
0169     app.SetParameterValue("unf:latency_us", 1'000'000 / 20'000); // 20 kHz
0170     app.SetParameterValue("pefac:latency_us", 1'000'000 / 5000); // 5 kHz
0171     app.SetParameterValue("peproc:latency_us", 1'000'000 / 20'000); // 20 kHz
0172     app.SetParameterValue("benchmark:resultsdir", "docs/perf_tests");
0173     app.SetParameterValue("benchmark:rates_filename", "unfold_small_saturation.dat");
0174     app.SetParameterValue("benchmark:use_log_scale", false);
0175     app.SetParameterValue("benchmark:minthreads", "1");
0176     app.SetParameterValue("benchmark:maxthreads", "16");
0177
0178     app.Add(new BSrc);
0179     app.Add(new Unf);
0180     app.Add(new PEProc);
0181     app.Add(new JFactoryGeneratorT<BFac>);
0182     app.Add(new JFactoryGeneratorT<PEFac>);
0183
0184     JBenchmarker benchmarker(&app);
0185     benchmarker.RunUntilFinished();
0186 }
0187
0188 TEST_CASE("UnfoldTopology_Medium") {
0189     LOG << "Running UnfoldTopology_Medium";
0190
0191     JApplication app;
0192     app.SetParameterValue("bsrc:latency_us", 0); // Infinity Hz
0193     app.SetParameterValue("bfac:latency_us", 1'000'000 / 10); // 10 Hz
0194     app.SetParameterValue("unf:latency_us", 0); // Infinity Hz
0195     app.SetParameterValue("pefac:latency_us", 1'000'000 / 100); // 100 Hz
0196     app.SetParameterValue("peproc:latency_us", 0); // Infinity Hz
0197     app.SetParameterValue("benchmark:resultsdir", "docs/perf_tests");
0198     app.SetParameterValue("benchmark:rates_filename", "unfold_medium.dat");
0199     app.SetParameterValue("benchmark:use_log_scale", true);
0200     app.SetParameterValue("benchmark:minthreads", "1");
0201     app.SetParameterValue("benchmark:maxthreads", "32");
0202
0203     app.Add(new BSrc);
0204     app.Add(new Unf);
0205     app.Add(new PEProc);
0206     app.Add(new JFactoryGeneratorT<BFac>);
0207     app.Add(new JFactoryGeneratorT<PEFac>);
0208
0209     JBenchmarker benchmarker(&app);
0210     benchmarker.RunUntilFinished();
0211 }
0212
0213 TEST_CASE("UnfoldTopology_Medium_Saturation") {
0214     LOG << "Running UnfoldTopology_Medium";
0215
0216     JApplication app;
0217     app.SetParameterValue("bsrc:latency_us", 1'000'000 / 40); // 40 Hz
0218     app.SetParameterValue("bfac:latency_us", 1'000'000 / 10); // 10 Hz
0219     app.SetParameterValue("unf:latency_us", 1'000'000 / 400); // 400 Hz
0220     app.SetParameterValue("pefac:latency_us", 1'000'000 / 100); // 100 Hz
0221     app.SetParameterValue("peproc:latency_us", 1'000'000 / 400); // 400 Hz
0222     app.SetParameterValue("benchmark:resultsdir", "docs/perf_tests");
0223     app.SetParameterValue("benchmark:rates_filename", "unfold_medium_saturation.dat");
0224     app.SetParameterValue("benchmark:use_log_scale", false);
0225     app.SetParameterValue("benchmark:minthreads", "1");
0226     app.SetParameterValue("benchmark:maxthreads", "16");
0227
0228     app.Add(new BSrc);
0229     app.Add(new Unf);
0230     app.Add(new PEProc);
0231     app.Add(new JFactoryGeneratorT<BFac>);
0232     app.Add(new JFactoryGeneratorT<PEFac>);
0233
0234     JBenchmarker benchmarker(&app);
0235     benchmarker.RunUntilFinished();
0236 }
0237
0238 TEST_CASE("UnfoldTopology_Large") {
0239     LOG << "Running UnfoldTopology_Large";
0240
0241     JApplication app;
0242     app.SetParameterValue("bsrc:latency_us", 0); // Infinity Hz
0243     app.SetParameterValue("bfac:latency_us", 1'000'000*2); // 0.5 Hz
0244     app.SetParameterValue("unf:latency_us", 0); // Infinity Hz
0245     app.SetParameterValue("pefac:latency_us", 1'000'000 / 5); // 5 Hz
0246     app.SetParameterValue("peproc:latency_us", 0); // Infinity Hz
0247     app.SetParameterValue("benchmark:resultsdir", "docs/perf_tests");
0248     app.SetParameterValue("benchmark:rates_filename", "unfold_large.dat");
0249     app.SetParameterValue("benchmark:use_log_scale", true);
0250     app.SetParameterValue("benchmark:minthreads", "1");
0251     app.SetParameterValue("benchmark:maxthreads", "32");
0252
0253     app.Add(new BSrc);
0254     app.Add(new Unf);
0255     app.Add(new PEProc);
0256     app.Add(new JFactoryGeneratorT<BFac>);
0257     app.Add(new JFactoryGeneratorT<PEFac>);
0258
0259     JBenchmarker benchmarker(&app);
0260     benchmarker.RunUntilFinished();
0261 }
0262
0263
0264 TEST_CASE("UnfoldTopology_Large_Saturation") {
0265     LOG << "Running UnfoldTopology_Large_Saturation";
0266
0267     JApplication app;
0268     app.SetParameterValue("bsrc:latency_us", 1'000'000 / 2); // 2 Hz
0269     app.SetParameterValue("bfac:latency_us", 1'000'000*2); // 0.5 Hz
0270     app.SetParameterValue("unf:latency_us", 1'000'000 / 20); // 20 Hz
0271     app.SetParameterValue("pefac:latency_us", 1'000'000 / 5); // 5 Hz
0272     app.SetParameterValue("peproc:latency_us", 1'000'000 / 20); // 20 Hz
0273     app.SetParameterValue("benchmark:resultsdir", "docs/perf_tests");
0274     app.SetParameterValue("benchmark:rates_filename", "unfold_large_saturation.dat");
0275     app.SetParameterValue("benchmark:use_log_scale", false);
0276     app.SetParameterValue("benchmark:minthreads", "1");
0277     app.SetParameterValue("benchmark:maxthreads", "16");
0278
0279     app.Add(new BSrc);
0280     app.Add(new Unf);
0281     app.Add(new PEProc);
0282     app.Add(new JFactoryGeneratorT<BFac>);
0283     app.Add(new JFactoryGeneratorT<PEFac>);
0284
0285     JBenchmarker benchmarker(&app);
0286     benchmarker.RunUntilFinished();
0287 }
0288
0289
0290 }
0291
0292
0293