Merge bitcoin#23025: bench: update nanobench add -min_time

e148a52 bench: fixed ubsan implicit conversion (Martin Ankerl)
da4e2f1 bench: various args improvements (Jon Atack)
d312fd9 bench: clean up includes (Jon Atack)
1f10f16 bench: add usage description and documentation (Martin Ankerl)
d3c6f8b bench: introduce -min_time argument (Martin Ankerl)
9fef832 bench: make EvictionProtection.* work with any number of iterations (Martin Ankerl)
153e686 bench: change AddrManGood to AddrManAddThenGood (Martin Ankerl)
468b232 bench: remove unnecessary & incorrect  multiplication in MuHashDiv (Martin Ankerl)
eed99cf bench: update nanobench from 4.3.4 to 4.3.6 (Martin Ankerl)

Pull request description:

  This PR updates the nanobench with the latest release from upstream, v4.3.6. It fixes the missing performance counters.

  Due to discussions on bitcoin#22999 I have done some work that should make the benchmark results more reliable. It introduces a new flag `-min_time` that allows to run a benchmark for much longer then the default. When results are unreliable, choosing a large timeframe here should usually get repeatable results even when frequency scaling cannot be disabled. The default is now 10ms. For this to work I have changed the `AddrManGood` and `EvictionProtection` benchmarks so they work with any number of iterations.

  Also, this adds more usage documentation to `bench_bitcoin -h` and I've cherry-picked two changes from bitcoin#22999 authored by Jon Atack

ACKs for top commit:
  jonatack:
    re-ACK e148a52
  laanwj:
    Code review ACK e148a52

Tree-SHA512: 2da6de19a5c85ac234b190025e195c727546166dbb75e3f9267e667a73677ba1e29b7765877418a42b1407b65df901e0130763936525e6f1450f18f08837c40c

Author: laanwj

src/bench/bench.cpp

@@ -4,15 +4,20 @@
 
 #include <bench/bench.h>
 
-#include <chainparams.h>
 #include <fs.h>
 #include <test/util/setup_common.h>
-#include <validation.h>
 
 #include <algorithm>
+#include <chrono>
 #include <fstream>
+#include <functional>
 #include <iostream>
+#include <map>
 #include <regex>
+#include <string>
+#include <vector>
+
+using namespace std::chrono_literals;
 
 const std::function<void(const std::string&)> G_TEST_LOG_FUN{};
 
@@ -66,6 +71,12 @@ void benchmark::BenchRunner::RunAll(const Args& args)
 
         Bench bench;
         bench.name(p.first);
+        if (args.min_time > 0ms) {
+            // convert to nanos before dividing to reduce rounding errors
+            std::chrono::nanoseconds min_time_ns = args.min_time;
+            bench.minEpochTime(min_time_ns / bench.epochs());
+        }
+
         if (args.asymptote.empty()) {
             p.second(bench);
         } else {

src/bench/bench.h

@@ -42,11 +42,12 @@ using ankerl::nanobench::Bench;
 typedef std::function<void(Bench&)> BenchFunction;
 
 struct Args {
-    std::string regex_filter;
     bool is_list_only;
+    std::chrono::milliseconds min_time;
     std::vector<double> asymptote;
     fs::path output_csv;
     fs::path output_json;
+    std::string regex_filter;
 };
 
 class BenchRunner

src/bench/bench_bitcoin.cpp

@@ -4,23 +4,31 @@
 
 #include <bench/bench.h>
 
+#include <clientversion.h>
 #include <crypto/sha256.h>
 #include <fs.h>
 #include <stacktraces.h>
 #include <util/strencodings.h>
 #include <util/system.h>
 
 #include <bls/bls.h>
+#include <chrono>
+#include <cstdint>
+#include <iostream>
+#include <sstream>
+#include <vector>
 
 static const char* DEFAULT_BENCH_FILTER = ".*";
+static constexpr int64_t DEFAULT_MIN_TIME_MS{10};
 
 static void SetupBenchArgs(ArgsManager& argsman)
 {
     SetupHelpOptions(argsman);
 
-    argsman.AddArg("-asymptote=n1,n2,n3,...", "Test asymptotic growth of the runtime of an algorithm, if supported by the benchmark", ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
+    argsman.AddArg("-asymptote=<n1,n2,n3,...>", "Test asymptotic growth of the runtime of an algorithm, if supported by the benchmark", ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
     argsman.AddArg("-filter=<regex>", strprintf("Regular expression filter to select benchmark by name (default: %s)", DEFAULT_BENCH_FILTER), ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
-    argsman.AddArg("-list", "List benchmarks without executing them", ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
+    argsman.AddArg("-list", "List benchmarks without executing them", ArgsManager::ALLOW_BOOL, OptionsCategory::OPTIONS);
+    argsman.AddArg("-min_time=<milliseconds>", strprintf("Minimum runtime per benchmark, in milliseconds (default: %d)", DEFAULT_MIN_TIME_MS), ArgsManager::ALLOW_INT, OptionsCategory::OPTIONS);
     argsman.AddArg("-output_csv=<output.csv>", "Generate CSV file with the most important benchmark results", ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
     argsman.AddArg("-output_json=<output.json>", "Generate JSON file with all benchmark results", ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
 }
@@ -50,16 +58,62 @@ int main(int argc, char** argv)
     }
 
     if (HelpRequested(argsman)) {
-        std::cout << argsman.GetHelpMessage();
+        std::cout << "Usage:  bench_dash [options]\n"
+                     "\n"
+                  << argsman.GetHelpMessage()
+                  << "Description:\n"
+                     "\n"
+                     "  bench_dash executes microbenchmarks. The quality of the benchmark results\n"
+                     "  highly depend on the stability of the machine. It can sometimes be difficult\n"
+                     "  to get stable, repeatable results, so here are a few tips:\n"
+                     "\n"
+                     "  * Use pyperf [1] to disable frequency scaling, turbo boost etc. For best\n"
+                     "    results, use CPU pinning and CPU isolation (see [2]).\n"
+                     "\n"
+                     "  * Each call of run() should do exactly the same work. E.g. inserting into\n"
+                     "    a std::vector doesn't do that as it will reallocate on certain calls. Make\n"
+                     "    sure each run has exactly the same preconditions.\n"
+                     "\n"
+                     "  * If results are still not reliable, increase runtime with e.g.\n"
+                     "    -min_time=5000 to let a benchmark run for at least 5 seconds.\n"
+                     "\n"
+                     "  * bench_dash uses nanobench [3] for which there is extensive\n"
+                     "    documentation available online.\n"
+                     "\n"
+                     "Environment Variables:\n"
+                     "\n"
+                     "  To attach a profiler you can run a benchmark in endless mode. This can be\n"
+                     "  done with the environment variable NANOBENCH_ENDLESS. E.g. like so:\n"
+                     "\n"
+                     "    NANOBENCH_ENDLESS=MuHash ./bench_dash -filter=MuHash\n"
+                     "\n"
+                     "  In rare cases it can be useful to suppress stability warnings. This can be\n"
+                     "  done with the environment variable NANOBENCH_SUPPRESS_WARNINGS, e.g:\n"
+                     "\n"
+                     "    NANOBENCH_SUPPRESS_WARNINGS=1 ./bench_dash\n"
+                     "\n"
+                     "Notes:\n"
+                     "\n"
+                     "  1. pyperf\n"
+                     "     https://github.com/psf/pyperf\n"
+                     "\n"
+                     "  2. CPU pinning & isolation\n"
+                     "     https://pyperf.readthedocs.io/en/latest/system.html\n"
+                     "\n"
+                     "  3. nanobench\n"
+                     "     https://github.com/martinus/nanobench\n"
+                     "\n";
+
         return EXIT_SUCCESS;
     }
 
     benchmark::Args args;
-    args.regex_filter = argsman.GetArg("-filter", DEFAULT_BENCH_FILTER);
-    args.is_list_only = argsman.GetBoolArg("-list", false);
     args.asymptote = parseAsymptote(argsman.GetArg("-asymptote", ""));
+    args.is_list_only = argsman.GetBoolArg("-list", false);
+    args.min_time = std::chrono::milliseconds(argsman.GetArg("-min_time", DEFAULT_MIN_TIME_MS));
     args.output_csv = fs::PathFromString(argsman.GetArg("-output_csv", ""));
     args.output_json = fs::PathFromString(argsman.GetArg("-output_json", ""));
+    args.regex_filter = argsman.GetArg("-filter", DEFAULT_BENCH_FILTER);
 
     benchmark::BenchRunner::RunAll(args);
 

src/bench/crypto_hash.cpp

@@ -249,9 +249,9 @@ static void MuHash(benchmark::Bench& bench)
 {
     MuHash3072 acc;
     unsigned char key[32] = {0};
-    int i = 0;
+    uint32_t i = 0;
     bench.run([&] {
-        key[0] = ++i;
+        key[0] = ++i & 0xFF;
         acc *= MuHash3072(key);
     });
 }
@@ -273,10 +273,6 @@ static void MuHashDiv(benchmark::Bench& bench)
     FastRandomContext rng(true);
     MuHash3072 muhash{rng.randbytes(32)};
 
-    for (size_t i = 0; i < bench.epochIterations(); ++i) {
-        acc *= muhash;
-    }
-
     bench.run([&] {
         acc /= muhash;
     });

src/bench/peer_eviction.cpp

@@ -20,19 +20,17 @@ static void EvictionProtectionCommon(
 {
     using Candidates = std::vector<NodeEvictionCandidate>;
     FastRandomContext random_context{true};
-    bench.warmup(100).epochIterations(1100);
 
     Candidates candidates{GetRandomNodeEvictionCandidates(num_candidates, random_context)};
     for (auto& c : candidates) {
         candidate_setup_fn(c);
     }
 
-    std::vector<Candidates> copies{
-        static_cast<size_t>(bench.epochs() * bench.epochIterations()), candidates};
-    size_t i{0};
+
     bench.run([&] {
-        ProtectEvictionCandidatesByRatio(copies.at(i));
-        ++i;
+        // creating a copy has an overhead of about 3%, so it does not influence the benchmark results much.
+        auto copy = candidates;
+        ProtectEvictionCandidatesByRatio(copy);
     });
 }
 

src/bench/rollingbloom.cpp

@@ -13,16 +13,16 @@ static void RollingBloom(benchmark::Bench& bench)
     uint32_t count = 0;
     bench.run([&] {
         count++;
-        data[0] = count;
-        data[1] = count >> 8;
-        data[2] = count >> 16;
-        data[3] = count >> 24;
+        data[0] = count & 0xFF;
+        data[1] = (count >> 8) & 0xFF;
+        data[2] = (count >> 16) & 0xFF;
+        data[3] = (count >> 24) & 0xFF;
         filter.insert(data);
 
-        data[0] = count >> 24;
-        data[1] = count >> 16;
-        data[2] = count >> 8;
-        data[3] = count;
+        data[0] = (count >> 24) & 0xFF;
+        data[1] = (count >> 16) & 0xFF;
+        data[2] = (count >> 8) & 0xFF;
+        data[3] = count & 0xFF;
         filter.contains(data);
     });
 }