Print metrics for blackbox tests

It was overcounting

Bench.cpp

@@ -358,7 +358,30 @@ void benchWorstCaseForRadixRangeRead() {
   // }
 }
 
+void benchMetrics() {
+  ankerl::nanobench::Bench bench;
+  ConflictSet cs{0};
+
+  int count;
+  ConflictSet::MetricsV1 *m;
+  cs.getMetricsV1(&m, &count);
+  bench.batch(count);
+  bench.run("fetch metric", [&]() {
+    for (int i = 0; i < count; ++i) {
+      m[i].getValue();
+    }
+  });
+}
+
+void benchCreateAndDestroy() {
+  ankerl::nanobench::Bench bench;
+
+  bench.run("create and destroy", [&]() { ConflictSet cs{0}; });
+}
+
 int main(void) {
   benchConflictSet();
   benchWorstCaseForRadixRangeRead();
+  benchMetrics();
+  benchCreateAndDestroy();
 }

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.18)
 project(
   conflict-set
-  VERSION 0.0.7
+  VERSION 0.0.9
   DESCRIPTION
     "A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys."
   HOMEPAGE_URL "https://git.weaselab.dev/weaselab/conflict-set"
@@ -60,6 +60,8 @@ cmake_pop_check_state()
 option(USE_SIMD_FALLBACK
        "Use fallback implementations of functions that use SIMD" OFF)
 
+option(DISABLE_TSAN "Disable TSAN" OFF)
+
 # This is encouraged according to
 # https://valgrind.org/docs/manual/manual-core-adv.html#manual-core-adv.clientreq
 include_directories(SYSTEM ${CMAKE_CURRENT_SOURCE_DIR}/third_party/valgrind)
@@ -245,7 +247,7 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
   endforeach()
 
   # tsan tests
-  if(NOT CMAKE_CROSSCOMPILING AND NOT CMAKE_BUILD_TYPE STREQUAL Debug)
+  if(NOT CMAKE_CROSSCOMPILING AND NOT DISABLE_TSAN)
     add_executable(tsan_driver ConflictSet.cpp FuzzTestDriver.cpp)
     target_compile_options(tsan_driver PRIVATE ${TEST_FLAGS} -fsanitize=thread)
     target_link_options(tsan_driver PRIVATE -fsanitize=thread)

ConflictSet.cpp

@@ -29,6 +29,7 @@ limitations under the License.
 #include <span>
 #include <string>
 #include <string_view>
+#include <sys/time.h>
 #include <type_traits>
 #include <utility>
 
@@ -546,7 +547,48 @@ static_assert(kBytesPerKey - sizeof(Node0) >= kMinNodeSurplus);
 
 constexpr int64_t kFreeListMaxMemory = 1 << 20;
 
+struct Metric {
+  Metric *prev;
+  const char *name;
+  const char *help;
+  ConflictSet::MetricsV1::Type type;
+  std::atomic<double> value;
+
+protected:
+  Metric(ConflictSet::Impl *impl, const char *name, const char *help,
+         ConflictSet::MetricsV1::Type type);
+};
+
+struct Gauge : private Metric {
+  Gauge(ConflictSet::Impl *impl, const char *name, const char *help)
+      : Metric(impl, name, help, ConflictSet::MetricsV1::Gauge) {}
+
+  void set(double value) {
+    this->value.store(value, std::memory_order_relaxed);
+  }
+};
+
+struct Counter : private Metric {
+  Counter(ConflictSet::Impl *impl, const char *name, const char *help)
+      : Metric(impl, name, help, ConflictSet::MetricsV1::Counter) {}
+  // Expensive. Accumulate locally and then call add instead of repeatedly
+  // calling add.
+  void add(double value) {
+    assert(value >= 0);
+    static_assert(std::atomic<double>::is_always_lock_free);
+    double old = this->value.load(std::memory_order_relaxed);
+    for (;;) {
+      double newVal = old + value;
+      if (this->value.compare_exchange_weak(old, newVal,
+                                            std::memory_order_relaxed)) {
+        break;
+      }
+    }
+  }
+};
+
 template <class T> struct BoundedFreeListAllocator {
+
   static_assert(sizeof(T) >= sizeof(void *));
   static_assert(std::derived_from<T, Node>);
   static_assert(std::is_trivial_v<T>);
@@ -1670,17 +1712,30 @@ struct SearchStepWise {
   }
 };
 
+thread_local double point_read_accum = 0;
+thread_local double prefix_read_accum = 0;
+thread_local double range_read_accum = 0;
+thread_local double point_read_short_circuit_accum = 0;
+thread_local double prefix_read_short_circuit_accum = 0;
+thread_local double range_read_short_circuit_accum = 0;
+thread_local double point_read_iterations_accum = 0;
+thread_local double prefix_read_iterations_accum = 0;
+thread_local double range_read_iterations_accum = 0;
+thread_local double range_read_node_scan_accum = 0;
+
 // Logically this is the same as performing firstGeq and then checking against
 // point or range version according to cmp, but this version short circuits as
 // soon as it can prove that there's no conflict.
 bool checkPointRead(Node *n, const std::span<const uint8_t> key,
                     InternalVersionT readVersion, ConflictSet::Impl *impl) {
+  ++point_read_accum;
 #if DEBUG_VERBOSE && !defined(NDEBUG)
   fprintf(stderr, "Check point read: %s\n", printable(key).c_str());
 #endif
   auto remaining = key;
-  for (;;) {
+  for (;; ++point_read_iterations_accum) {
     if (maxVersion(n, impl) <= readVersion) {
+      ++point_read_short_circuit_accum;
       return true;
     }
     if (remaining.size() == 0) {
@@ -1752,17 +1807,19 @@ downLeftSpine:
 // short circuits as soon as it can prove that there's no conflict.
 bool checkPrefixRead(Node *n, const std::span<const uint8_t> key,
                      InternalVersionT readVersion, ConflictSet::Impl *impl) {
+  ++prefix_read_accum;
 #if DEBUG_VERBOSE && !defined(NDEBUG)
   fprintf(stderr, "Check prefix read: %s\n", printable(key).c_str());
 #endif
   auto remaining = key;
-  for (;;) {
+  for (;; ++prefix_read_iterations_accum) {
     auto m = maxVersion(n, impl);
     if (remaining.size() == 0) {
       return m <= readVersion;
     }
 
     if (m <= readVersion) {
+      ++prefix_read_short_circuit_accum;
       return true;
     }
 
@@ -1993,6 +2050,7 @@ scan16(const InternalVersionT *vs, int begin, int end,
 template <bool kAVX512>
 bool checkMaxBetweenExclusive(Node *n, int begin, int end,
                               InternalVersionT readVersion) {
+  ++range_read_node_scan_accum;
   assume(-1 <= begin);
   assume(begin <= 256);
   assume(-1 <= end);
@@ -2601,13 +2659,16 @@ bool checkRangeReadImpl(Node *n, std::span<const uint8_t> begin,
     return checkPrefixRead(n, begin, readVersion, impl);
   }
 
+  ++range_read_accum;
+
   SearchStepWise search{n, begin.subspan(0, lcp)};
   Arena arena;
-  for (;;) {
+  for (;; ++range_read_iterations_accum) {
     assert(getSearchPath(arena, search.n) <=>
                begin.subspan(0, lcp - search.remaining.size()) ==
            0);
     if (maxVersion(search.n, impl) <= readVersion) {
+      ++range_read_short_circuit_accum;
       return true;
     }
     if (search.step()) {
@@ -2648,18 +2709,21 @@ bool checkRangeReadImpl(Node *n, std::span<const uint8_t> begin,
     bool leftDone = checkRangeLeftSide.step();
     bool rightDone = checkRangeRightSide.step();
     if (!leftDone && !rightDone) {
+      range_read_iterations_accum += 2;
       continue;
     }
     if (leftDone && rightDone) {
       break;
     } else if (leftDone) {
-      while (!checkRangeRightSide.step())
-        ;
+      while (!checkRangeRightSide.step()) {
+        ++range_read_iterations_accum;
+      }
       break;
     } else {
       assert(rightDone);
-      while (!checkRangeLeftSide.step())
-        ;
+      while (!checkRangeLeftSide.step()) {
+        ++range_read_iterations_accum;
+      }
     }
     break;
   }
@@ -3004,11 +3068,22 @@ Node *firstGeqPhysical(Node *n, const std::span<const uint8_t> key) {
   }
 }
 
+#define MEASURE_CHECK_CPU_TIME 0
+
 struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
 
   void check(const ReadRange *reads, Result *result, int count) {
+#if MEASURE_CHECK_CPU_TIME
+    timespec ts_begin;
+    clock_gettime(CLOCK_THREAD_CPUTIME_ID, &ts_begin);
+#endif
+    int commits_accum = 0;
+    int conflicts_accum = 0;
+    int too_olds_accum = 0;
+    double check_byte_accum = 0;
     for (int i = 0; i < count; ++i) {
       const auto &r = reads[i];
+      check_byte_accum += r.begin.len + r.end.len;
       auto begin = std::span<const uint8_t>(r.begin.p, r.begin.len);
       auto end = std::span<const uint8_t>(r.end.p, r.end.len);
       assert(oldestVersionFullPrecision >=
@@ -3022,7 +3097,38 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
                                   InternalVersionT(reads[i].readVersion), this))
               ? Commit
               : Conflict;
+      commits_accum += result[i] == Commit;
+      conflicts_accum += result[i] == Conflict;
+      too_olds_accum += result[i] == TooOld;
     }
+    point_read_total.add(std::exchange(point_read_accum, 0));
+    prefix_read_total.add(std::exchange(prefix_read_accum, 0));
+    range_read_total.add(std::exchange(range_read_accum, 0));
+    range_read_node_scan_total.add(
+        std::exchange(range_read_node_scan_accum, 0));
+    point_read_short_circuit_total.add(
+        std::exchange(point_read_short_circuit_accum, 0));
+    prefix_read_short_circuit_total.add(
+        std::exchange(prefix_read_short_circuit_accum, 0));
+    range_read_short_circuit_total.add(
+        std::exchange(range_read_short_circuit_accum, 0));
+    point_read_iterations_total.add(
+        std::exchange(point_read_iterations_accum, 0));
+    prefix_read_iterations_total.add(
+        std::exchange(prefix_read_iterations_accum, 0));
+    range_read_iterations_total.add(
+        std::exchange(range_read_iterations_accum, 0));
+    commits_total.add(commits_accum);
+    conflicts_total.add(conflicts_accum);
+    too_olds_total.add(too_olds_accum);
+    check_bytes_total.add(check_byte_accum);
+#if MEASURE_CHECK_CPU_TIME
+    timespec ts_end;
+    clock_gettime(CLOCK_THREAD_CPUTIME_ID, &ts_end);
+    check_cpu_seconds_total.add(
+        std::max<double>(0, (ts_end.tv_nsec * 1e-9 + ts_end.tv_sec) -
+                                (ts_begin.tv_nsec * 1e-9 + ts_begin.tv_sec)));
+#endif
   }
 
   void addWrites(const WriteRange *writes, int count, int64_t writeVersion) {
@@ -3056,6 +3162,8 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
                       this);
       }
     }
+
+    memory_bytes.set(totalBytes);
   }
 
   // Spends up to `fuel` gc'ing, and returns its unused fuel. Reclaims memory
@@ -3070,8 +3178,12 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
       rootMaxVersion = std::max(rootMaxVersion, oldestVersion);
       n = nextPhysical(n);
     }
-    for (; fuel > 0 && n != nullptr; --fuel) {
+    for (; fuel > 0 && n != nullptr;) {
       rezero(n, oldestVersion);
+      // The "make sure gc keeps up with writes" calculations assume that we're
+      // scanning key by key, not node by node. Make sure we only spend fuel
+      // when there's a logical entry.
+      fuel -= n->entryPresent;
       if (n->entryPresent && std::max(n->entry.pointVersion,
                                       n->entry.rangeVersion) <= oldestVersion) {
         // Any transaction n would have prevented from committing is
@@ -3121,6 +3233,8 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     }
 #endif
     keyUpdates = gcScanStep(keyUpdates);
+
+    memory_bytes.set(totalBytes);
   }
 
   int64_t getBytes() const { return totalBytes; }
@@ -3156,8 +3270,14 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     // Intentionally not resetting totalBytes
   }
 
-  explicit Impl(int64_t oldestVersion) { init(oldestVersion); }
-  ~Impl() { destroyTree(root); }
+  explicit Impl(int64_t oldestVersion) {
+    init(oldestVersion);
+    initMetrics();
+  }
+  ~Impl() {
+    destroyTree(root);
+    safe_free(metrics, metricsCount * sizeof(metrics[0]));
+  }
 
   NodeAllocators allocators;
 
@@ -3173,8 +3293,77 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   int64_t oldestVersionAtGcBegin;
   int64_t newestVersionFullPrecision;
   int64_t totalBytes = 0;
+
+  MetricsV1 *metrics;
+  int metricsCount = 0;
+  void initMetrics() {
+    metrics = (MetricsV1 *)safe_malloc(metricsCount * sizeof(metrics[0]));
+    for (auto [i, m] = std::make_tuple(metricsCount - 1, metricList); i >= 0;
+         --i, m = m->prev) {
+      metrics[i].name = m->name;
+      metrics[i].help = m->help;
+      metrics[i].p = m;
+      metrics[i].type = m->type;
+    }
+  }
+
+  Metric *metricList = nullptr;
+
+#define GAUGE(name, help)                                                      \
+  Gauge name { this, #name, help }
+#define COUNTER(name, help)                                                    \
+  Counter name { this, #name, help }
+  // ==================== METRICS DEFINITIONS ====================
+  GAUGE(memory_bytes, "Total number of bytes in use");
+  COUNTER(point_read_total, "Total number of point reads checked");
+  COUNTER(point_read_short_circuit_total,
+          "Total number of point reads that did not require a full search to "
+          "check");
+  COUNTER(point_read_iterations_total,
+          "Total number of iterations of the main loop for point read checks");
+  COUNTER(prefix_read_total, "Total number of prefix reads checked");
+  COUNTER(prefix_read_short_circuit_total,
+          "Total number of prefix reads that did not require a full search to "
+          "check");
+  COUNTER(prefix_read_iterations_total,
+          "Total number of iterations of the main loop for prefix read checks");
+  COUNTER(range_read_total, "Total number of range reads checked");
+  COUNTER(range_read_short_circuit_total,
+          "Total number of range reads that did not require a full search to "
+          "check");
+  COUNTER(range_read_iterations_total,
+          "Total number of iterations of the main loops for range read checks");
+  COUNTER(range_read_node_scan_total,
+          "Total number of scans of individual nodes while "
+          "checking a range read");
+  COUNTER(commits_total,
+          "Total number of checks where the result is \"commit\"");
+  COUNTER(conflicts_total,
+          "Total number of checks where the result is \"conflict\"");
+  COUNTER(too_olds_total,
+          "Total number of checks where the result is \"too old\"");
+  COUNTER(check_bytes_total, "Total number of key bytes checked");
+#if MEASURE_CHECK_CPU_TIME
+  COUNTER(check_cpu_seconds_total,
+          "Total cpu seconds spent in a call to check");
+#endif
+  // ==================== END METRICS DEFINITIONS ====================
+#undef GAUGE
+#undef COUNTER
+
+  void getMetricsV1(MetricsV1 **metrics, int *count) {
+    *metrics = this->metrics;
+    *count = metricsCount;
+  }
 };
 
+Metric::Metric(ConflictSet::Impl *impl, const char *name, const char *help,
+               ConflictSet::MetricsV1::Type type)
+    : prev(std::exchange(impl->metricList, this)), name(name), help(help),
+      type(type), value(0) {
+  ++impl->metricsCount;
+}
+
 InternalVersionT maxVersion(Node *n, ConflictSet::Impl *impl) {
   int index = n->parentsIndex;
   n = n->parent;
@@ -3304,6 +3493,15 @@ void internal_destroy(ConflictSet::Impl *impl) {
 
 int64_t internal_getBytes(ConflictSet::Impl *impl) { return impl->getBytes(); }
 
+void internal_getMetricsV1(ConflictSet::Impl *impl,
+                           ConflictSet::MetricsV1 **metrics, int *count) {
+  impl->getMetricsV1(metrics, count);
+}
+
+double internal_getMetricValue(const ConflictSet::MetricsV1 *metric) {
+  return ((Metric *)metric->p)->value.load(std::memory_order_relaxed);
+}
+
 // ==================== END IMPLEMENTATION ====================
 
 // GCOVR_EXCL_START
@@ -3388,6 +3586,14 @@ void ConflictSet::setOldestVersion(int64_t oldestVersion) {
 
 int64_t ConflictSet::getBytes() const { return internal_getBytes(impl); }
 
+void ConflictSet::getMetricsV1(MetricsV1 **metrics, int *count) const {
+  return internal_getMetricsV1(impl, metrics, count);
+}
+
+double ConflictSet::MetricsV1::getValue() const {
+  return internal_getMetricValue(this);
+}
+
 ConflictSet::ConflictSet(int64_t oldestVersion)
     : impl(internal_create(oldestVersion)) {}
 
@@ -3443,6 +3649,7 @@ static_assert(std::is_standard_layout_v<ConflictSet::Result>);
 static_assert(std::is_standard_layout_v<ConflictSet::Key>);
 static_assert(std::is_standard_layout_v<ConflictSet::ReadRange>);
 static_assert(std::is_standard_layout_v<ConflictSet::WriteRange>);
+static_assert(std::is_standard_layout_v<ConflictSet::MetricsV1>);
 
 namespace {
 

Internal.h

@@ -664,17 +664,74 @@ template <class ConflictSetImpl> struct TestDriver {
       fprintf(stderr, "Write @ %" PRId64 "\n", v);
 #endif
 
+      // Test non-canonical writes
+      if (numPointWrites > 0) {
+        int overlaps = arbitrary.bounded(numPointWrites);
+        for (int i = 0; i < numPointWrites + numRangeWrites && overlaps > 0;
+             ++i) {
+          if (writes[i].end.len == 0) {
+            int keyLen = prefixLen + arbitrary.bounded(kMaxKeySuffixLen);
+            auto *begin = new (arena) uint8_t[keyLen];
+            memset(begin, prefixByte, prefixLen);
+            arbitrary.randomBytes(begin + prefixLen, keyLen - prefixLen);
+            writes[i].end.len = keyLen;
+            writes[i].end.p = begin;
+            auto c =
+                std::span<const uint8_t>(writes[i].begin.p,
+                                         writes[i].begin.len) <=>
+                std::span<const uint8_t>(writes[i].end.p, writes[i].end.len);
+            if (c > 0) {
+              using std::swap;
+              swap(writes[i].begin, writes[i].end);
+            } else if (c == 0) {
+              // It's a point write after all, I guess
+              writes[i].end.len = 0;
+            }
+            --overlaps;
+          }
+        }
+      }
+      if (arbitrary.bounded(2)) {
+        // Shuffle writes
+        for (int i = numPointWrites + numRangeWrites - 1; i > 0; --i) {
+          int j = arbitrary.bounded(i + 1);
+          if (i != j) {
+            using std::swap;
+            swap(writes[i], writes[j]);
+          }
+        }
+      }
+
+      oldestVersion +=
+          arbitrary.bounded(10) ? arbitrary.bounded(10) : arbitrary.next();
+      oldestVersion = std::min(oldestVersion, writeVersion);
+
+#ifdef THREAD_TEST
+      std::latch ready{1};
+      std::thread thread2{[&]() {
+        ready.count_down();
+        ConflictSet::MetricsV1 *m;
+        int count;
+        cs.getMetricsV1(&m, &count);
+        for (int i = 0; i < count; ++i) {
+          m[i].getValue();
+        }
+      }};
+      ready.wait();
+#endif
+
       CALLGRIND_START_INSTRUMENTATION;
       cs.addWrites(writes, numPointWrites + numRangeWrites, v);
       CALLGRIND_STOP_INSTRUMENTATION;
 
       refImpl.addWrites(writes, numPointWrites + numRangeWrites, v);
 
-      oldestVersion +=
-          arbitrary.bounded(10) ? arbitrary.bounded(10) : arbitrary.next();
-      oldestVersion = std::min(oldestVersion, writeVersion);
       cs.setOldestVersion(oldestVersion);
       refImpl.setOldestVersion(oldestVersion);
+
+#ifdef THREAD_TEST
+      thread2.join();
+#endif
     }
     {
       int numPointReads = arbitrary.bounded(100);
@@ -747,7 +804,15 @@ template <class ConflictSetImpl> struct TestDriver {
       std::latch ready{1};
       std::thread thread2{[&]() {
         ready.count_down();
+        // Call all const methods
         cs.check(reads, results3, numPointReads + numRangeReads);
+        cs.getBytes();
+        ConflictSet::MetricsV1 *m;
+        int count;
+        cs.getMetricsV1(&m, &count);
+        for (int i = 0; i < count; ++i) {
+          m[i].getValue();
+        }
       }};
       ready.wait();
 #endif
@@ -756,6 +821,15 @@ template <class ConflictSetImpl> struct TestDriver {
       cs.check(reads, results1, numPointReads + numRangeReads);
       CALLGRIND_STOP_INSTRUMENTATION;
 
+      // Call remaining const methods
+      cs.getBytes();
+      ConflictSet::MetricsV1 *m;
+      int count;
+      cs.getMetricsV1(&m, &count);
+      for (int i = 0; i < count; ++i) {
+        m[i].getValue();
+      }
+
       refImpl.check(reads, results2, numPointReads + numRangeReads);
 
       auto compareResults = [reads](ConflictSet::Result *results1,

Jenkinsfile

@@ -48,6 +48,17 @@ pipeline {
                 recordIssues(tools: [clang()])
             }
         }
+        stage('Debug') {
+            agent {
+                dockerfile {
+                    args '-v /home/jenkins/ccache:/ccache'
+                    reuseNode true
+                }
+            }
+            steps {
+                CleanBuildAndTest("-DCMAKE_BUILD_TYPE=Debug")
+            }
+        }
         stage('SIMD fallback') {
             agent {
                 dockerfile {
@@ -59,17 +70,6 @@ pipeline {
                 CleanBuildAndTest("-DUSE_SIMD_FALLBACK=ON")
             }
         }
-        stage('32-bit versions') {
-            agent {
-                dockerfile {
-                    args '-v /home/jenkins/ccache:/ccache'
-                    reuseNode true
-                }
-            }
-            steps {
-                CleanBuildAndTest("-DUSE_32_BIT_VERSIONS=ON")
-            }
-        }
         stage('Release [gcc]') {
             agent {
                 dockerfile {
@@ -117,7 +117,7 @@ pipeline {
                 }
             }
             steps {
-                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_C_FLAGS=--coverage -DCMAKE_CXX_FLAGS=--coverage -DCMAKE_BUILD_TYPE=Debug")
+                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_C_FLAGS=--coverage -DCMAKE_CXX_FLAGS=--coverage -DCMAKE_BUILD_TYPE=Debug -DDISABLE_TSAN=ON")
                 sh '''
                 gcovr -f ConflictSet.cpp --cobertura > build/coverage.xml
                 '''

README.md

@@ -60,27 +60,27 @@ Performance counters:
 
 |               ns/op |                op/s |    err% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------:|:----------
-|              256.89 |        3,892,784.92 |    0.3% |      0.01 | `point reads`
-|              272.90 |        3,664,395.04 |    0.2% |      0.01 | `prefix reads`
-|              507.22 |        1,971,549.50 |    0.7% |      0.01 | `range reads`
-|              452.66 |        2,209,181.91 |    0.5% |      0.01 | `point writes`
-|              438.09 |        2,282,619.96 |    0.4% |      0.01 | `prefix writes`
-|              253.33 |        3,947,420.36 |    2.5% |      0.02 | `range writes`
-|              574.07 |        1,741,936.71 |    0.3% |      0.01 | `monotonic increasing point writes`
-|          151,562.50 |            6,597.94 |    1.5% |      0.01 | `worst case for radix tree`
+|              245.99 |        4,065,232.81 |    0.3% |      0.01 | `point reads`
+|              265.93 |        3,760,430.49 |    0.2% |      0.01 | `prefix reads`
+|              485.30 |        2,060,569.50 |    0.2% |      0.01 | `range reads`
+|              449.60 |        2,224,195.17 |    0.4% |      0.01 | `point writes`
+|              441.76 |        2,263,688.18 |    1.1% |      0.01 | `prefix writes`
+|              245.42 |        4,074,647.54 |    2.4% |      0.02 | `range writes`
+|              572.80 |        1,745,810.06 |    1.3% |      0.01 | `monotonic increasing point writes`
+|          154,819.33 |            6,459.14 |    0.9% |      0.01 | `worst case for radix tree`
 
 ## Radix tree (this implementation)
 
 |               ns/op |                op/s |    err% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------:|:----------
-|               19.83 |       50,420,955.28 |    0.1% |      0.01 | `point reads`
-|               55.95 |       17,872,542.40 |    0.5% |      0.01 | `prefix reads`
-|               88.28 |       11,327,709.50 |    0.4% |      0.01 | `range reads`
-|               29.15 |       34,309,531.64 |    0.5% |      0.01 | `point writes`
-|               42.36 |       23,607,424.27 |    1.1% |      0.01 | `prefix writes`
-|               50.00 |       20,000,000.00 |    0.0% |      0.01 | `range writes`
-|               93.52 |       10,692,413.79 |    3.3% |      0.01 | `monotonic increasing point writes`
-|        2,388,417.00 |              418.69 |    0.4% |      0.03 | `worst case for radix tree`
+|               19.17 |       52,163,930.66 |    0.1% |      0.01 | `point reads`
+|               23.68 |       42,224,388.21 |    0.7% |      0.01 | `prefix reads`
+|               63.30 |       15,797,506.06 |    0.9% |      0.01 | `range reads`
+|               29.66 |       33,720,994.74 |    0.3% |      0.01 | `point writes`
+|               43.50 |       22,987,781.25 |    1.0% |      0.01 | `prefix writes`
+|               50.00 |       20,000,000.00 |    0.8% |      0.01 | `range writes`
+|              103.25 |        9,684,786.47 |    2.9% |      0.01 | `monotonic increasing point writes`
+|        1,181,500.00 |              846.38 |    2.3% |      0.01 | `worst case for radix tree`
 
 # "Real data" test
 

RealDataBench.cpp

@@ -129,6 +129,16 @@ int main(int argc, const char **argv) {
     close(fd);
   }
 
+  ConflictSet::MetricsV1 *metrics;
+  int metricsCount;
+  cs.getMetricsV1(&metrics, &metricsCount);
+  for (int i = 0; i < metricsCount; ++i) {
+    printf("# HELP %s %s\n", metrics[i].name, metrics[i].help);
+    printf("# TYPE %s %s\n", metrics[i].name,
+           metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
+    printf("%s %g\n", metrics[i].name, metrics[i].getValue());
+  }
+
   printf("Check: %g seconds, %g MB/s, Add: %g seconds, %g MB/s, Gc ratio: "
          "%g%%, Peak idle memory: %g\n",
          checkTime, checkBytes / checkTime * 1e-6, addTime,

SkipList.cpp

@@ -22,6 +22,8 @@
 
 #include "ConflictSet.h"
 #include "Internal.h"
+
+#include <algorithm>
 #include <span>
 
 std::span<const uint8_t> keyAfter(Arena &arena, std::span<const uint8_t> key) {
@@ -40,7 +42,7 @@ std::span<const uint8_t> copyToArena(Arena &arena,
 }
 
 using Version = int64_t;
-#define force_inline __attribute__((always_inline))
+#define force_inline inline __attribute__((always_inline))
 using StringRef = std::span<const uint8_t>;
 
 struct KeyRangeRef {
@@ -52,6 +54,135 @@ struct KeyRangeRef {
       : begin(begin), end(keyAfter(arena, begin)) {}
 };
 
+struct KeyInfo {
+  StringRef key;
+  bool begin;
+  bool write;
+
+  KeyInfo() = default;
+  KeyInfo(StringRef key, bool begin, bool write)
+      : key(key), begin(begin), write(write) {}
+};
+
+force_inline int extra_ordering(const KeyInfo &ki) {
+  return ki.begin * 2 + (ki.write ^ ki.begin);
+}
+
+// returns true if done with string
+force_inline bool getCharacter(const KeyInfo &ki, int character,
+                               int &outputCharacter) {
+  // normal case
+  if (character < ki.key.size()) {
+    outputCharacter = 5 + ki.key.begin()[character];
+    return false;
+  }
+
+  // termination
+  if (character == ki.key.size()) {
+    outputCharacter = 0;
+    return false;
+  }
+
+  if (character == ki.key.size() + 1) {
+    // end/begin+read/write relative sorting
+    outputCharacter = extra_ordering(ki);
+    return false;
+  }
+
+  outputCharacter = 0;
+  return true;
+}
+
+bool operator<(const KeyInfo &lhs, const KeyInfo &rhs) {
+  int i = std::min(lhs.key.size(), rhs.key.size());
+  int c = memcmp(lhs.key.data(), rhs.key.data(), i);
+  if (c != 0)
+    return c < 0;
+
+  // Always sort shorter keys before longer keys.
+  if (lhs.key.size() < rhs.key.size()) {
+    return true;
+  }
+  if (lhs.key.size() > rhs.key.size()) {
+    return false;
+  }
+
+  // When the keys are the same length, use the extra ordering constraint.
+  return extra_ordering(lhs) < extra_ordering(rhs);
+}
+
+bool operator==(const KeyInfo &lhs, const KeyInfo &rhs) {
+  return !(lhs < rhs || rhs < lhs);
+}
+
+void swapSort(std::vector<KeyInfo> &points, int a, int b) {
+  if (points[b] < points[a]) {
+    KeyInfo temp;
+    temp = points[a];
+    points[a] = points[b];
+    points[b] = temp;
+  }
+}
+
+struct SortTask {
+  int begin;
+  int size;
+  int character;
+  SortTask(int begin, int size, int character)
+      : begin(begin), size(size), character(character) {}
+};
+
+void sortPoints(std::vector<KeyInfo> &points) {
+  std::vector<SortTask> tasks;
+  std::vector<KeyInfo> newPoints;
+  std::vector<int> counts;
+
+  tasks.emplace_back(0, points.size(), 0);
+
+  while (tasks.size()) {
+    SortTask st = tasks.back();
+    tasks.pop_back();
+
+    if (st.size < 10) {
+      std::sort(points.begin() + st.begin, points.begin() + st.begin + st.size);
+      continue;
+    }
+
+    newPoints.resize(st.size);
+    counts.assign(256 + 5, 0);
+
+    // get counts
+    int c;
+    bool allDone = true;
+    for (int i = st.begin; i < st.begin + st.size; i++) {
+      allDone &= getCharacter(points[i], st.character, c);
+      counts[c]++;
+    }
+    if (allDone)
+      continue;
+
+    // calculate offsets from counts and build next level of tasks
+    int total = 0;
+    for (int i = 0; i < counts.size(); i++) {
+      int temp = counts[i];
+      if (temp > 1)
+        tasks.emplace_back(st.begin + total, temp, st.character + 1);
+      counts[i] = total;
+      total += temp;
+    }
+
+    // put in their places
+    for (int i = st.begin; i < st.begin + st.size; i++) {
+      getCharacter(points[i], st.character, c);
+      newPoints[counts[c]++] = points[i];
+    }
+
+    // copy back into original points array
+    for (int i = 0; i < st.size; i++)
+      points[st.begin + i] = newPoints[i];
+  }
+}
+
 static thread_local uint32_t g_seed = 0;
 
 static inline int skfastrand() {
@@ -602,10 +733,40 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
 
   void addWrites(const ConflictSet::WriteRange *writes, int count,
                  int64_t writeVersion) {
+    auto points = std::vector<KeyInfo>(count * 2);
+    Arena arena;
+
+    for (int r = 0; r < count; r++) {
+      points.emplace_back(StringRef(writes[r].begin.p, writes[r].begin.len),
+                          true, true);
+      points.emplace_back(
+          writes[r].end.len > 0
+              ? StringRef{writes[r].end.p, size_t(writes[r].end.len)}
+              : keyAfter(arena, points.back().key),
+          false, true);
+    }
+
+    sortPoints(points);
+
+    int activeWriteCount = 0;
+    std::vector<std::pair<StringRef, StringRef>> combinedWriteConflictRanges;
+    for (const KeyInfo &point : points) {
+      if (point.write) {
+        if (point.begin) {
+          activeWriteCount++;
+          if (activeWriteCount == 1)
+            combinedWriteConflictRanges.emplace_back(point.key, StringRef());
+        } else /*if (point.end)*/ {
+          activeWriteCount--;
+          if (activeWriteCount == 0)
+            combinedWriteConflictRanges.back().second = point.key;
+        }
+      }
+    }
+
     assert(writeVersion >= newestVersion);
     newestVersion = writeVersion;
-    Arena arena;
-    const int stringCount = count * 2;
+    const int stringCount = combinedWriteConflictRanges.size() * 2;
 
     const int stripeSize = 16;
     SkipList::Finger fingers[stripeSize];
@@ -616,15 +777,13 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     int ss = stringCount - (stripes - 1) * stripeSize;
     for (int s = stripes - 1; s >= 0; s--) {
       for (int i = 0; i * 2 < ss; ++i) {
-        const auto &w = writes[s * stripeSize / 2 + i];
+        const auto &w = combinedWriteConflictRanges[s * stripeSize / 2 + i];
 #if DEBUG_VERBOSE
         printf("Write begin: %s\n", printable(w.begin).c_str());
         fflush(stdout);
 #endif
-        values[i * 2] = {w.begin.p, size_t(w.begin.len)};
-        values[i * 2 + 1] = w.end.len > 0
-                                ? StringRef{w.end.p, size_t(w.end.len)}
-                                : keyAfter(arena, values[i * 2]);
+        values[i * 2] = w.first;
+        values[i * 2 + 1] = w.second;
         keyUpdates += 3;
       }
       skipList.find(values, fingers, temp, ss);
@@ -702,6 +861,16 @@ void internal_destroy(ConflictSet::Impl *impl) {
 
 int64_t internal_getBytes(ConflictSet::Impl *impl) { return impl->totalBytes; }
 
+void internal_getMetricsV1(ConflictSet::Impl *impl,
+                           ConflictSet::MetricsV1 **metrics, int *count) {
+  *metrics = nullptr;
+  *count = 0;
+}
+
+double internal_getMetricValue(const ConflictSet::MetricsV1 *metric) {
+  return 0;
+}
+
 void ConflictSet::check(const ReadRange *reads, Result *results,
                         int count) const {
   internal_check(impl, reads, results, count);
@@ -718,6 +887,14 @@ void ConflictSet::setOldestVersion(int64_t oldestVersion) {
 
 int64_t ConflictSet::getBytes() const { return internal_getBytes(impl); }
 
+void ConflictSet::getMetricsV1(MetricsV1 **metrics, int *count) const {
+  return internal_getMetricsV1(impl, metrics, count);
+}
+
+double ConflictSet::MetricsV1::getValue() const {
+  return internal_getMetricValue(this);
+}
+
 ConflictSet::ConflictSet(int64_t oldestVersion)
     : impl(internal_create(oldestVersion)) {}
 

TestDriver.cpp

@@ -15,5 +15,18 @@ int main(int argc, char **argv) {
     if (!driver.ok) {
       abort();
     }
+
+    ConflictSet::MetricsV1 *metrics;
+    int metricsCount;
+    driver.cs.getMetricsV1(&metrics, &metricsCount);
+    printf("#################### METRICS FOR %s ####################\n",
+           argv[i]);
+    for (int i = 0; i < metricsCount; ++i) {
+      printf("# HELP %s %s\n", metrics[i].name, metrics[i].help);
+      printf("# TYPE %s %s\n", metrics[i].name,
+             metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
+      printf("%s %g\n", metrics[i].name, metrics[i].getValue());
+    }
+    puts("");
   }
 }

aarch64-symbol-imports.txt

@@ -1,3 +1,5 @@
+__aarch64_cas8_relax
+__getauxval@GLIBC_2.17
 __stack_chk_fail@GLIBC_2.17
 __stack_chk_guard@GLIBC_2.17
 abort@GLIBC_2.17

apple-symbol-exports.txt

@@ -13,5 +13,7 @@ __ZN8weaselab11ConflictSetC2Ex
 __ZN8weaselab11ConflictSetD1Ev
 __ZN8weaselab11ConflictSetD2Ev
 __ZN8weaselab11ConflictSetaSEOS0_
+__ZNK8weaselab11ConflictSet12getMetricsV1EPPNS0_9MetricsV1EPi
 __ZNK8weaselab11ConflictSet5checkEPKNS0_9ReadRangeEPNS0_6ResultEi
 __ZNK8weaselab11ConflictSet8getBytesEv
+__ZNK8weaselab11ConflictSet9MetricsV18getValueEv

conflict_set_cxx_api_test.cpp

@@ -1,6 +1,7 @@
 #include "ConflictSet.h"
 
 #include <cassert>
+#include <cstdio>
 
 using namespace weaselab;
 
@@ -21,4 +22,14 @@ int main(void) {
   assert(result == ConflictSet::Conflict);
   int64_t bytes = cs.getBytes();
   assert(bytes > 0);
+
+  ConflictSet::MetricsV1 *metrics;
+  int metricsCount;
+  cs.getMetricsV1(&metrics, &metricsCount);
+  for (int i = 0; i < metricsCount; ++i) {
+    printf("# HELP %s %s\n", metrics[i].name, metrics[i].help);
+    printf("# TYPE %s %s\n", metrics[i].name,
+           metrics[i].type == metrics[i].Counter ? "counter" : "gauge");
+    printf("%s %g\n", metrics[i].name, metrics[i].getValue());
+  }
 }

corpus/0e410d4488defd877d2c9690a26970f57ab78b0c

@@ -1,4 +0,0 @@
-*
-
-
-

corpus/22c23c2418773edaaff055ed387f82c8ad078f9f

@@ -1,2 +0,0 @@
-*
-