AtomicHashMapTest.cpp

Go to the documentation of this file.

/*
 * Copyright 2012-present Facebook, Inc.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *   http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#include <folly/AtomicHashMap.h>

#include <atomic>
#include <memory>
#include <thread>

#include <glog/logging.h>

#include <folly/Benchmark.h>
#include <folly/Conv.h>
#include <folly/portability/GTest.h>
#include <folly/portability/SysTime.h>

using folly::AtomicHashArray;
using folly::AtomicHashMap;
using folly::StringPiece;
using std::string;
using std::vector;

// Tunables:
DEFINE_double(targetLoadFactor, 0.75, "Target memory utilization fraction.");
DEFINE_double(maxLoadFactor, 0.80, "Max before growth.");
DEFINE_int32(numThreads, 8, "Threads to use for concurrency tests.");
DEFINE_int64(numBMElements, 12 * 1000 * 1000, "Size of maps for benchmarks.");

const double LF = FLAGS_maxLoadFactor / FLAGS_targetLoadFactor;
const int maxBMElements = int(FLAGS_numBMElements * LF); // hit our target LF.

static int64_t nowInUsec() {
  timeval tv;
  gettimeofday(&tv, nullptr);
  return int64_t(tv.tv_sec) * 1000 * 1000 + tv.tv_usec;
}

TEST(Ahm, BasicStrings) {
  typedef AtomicHashMap<int64_t, string> AHM;
  AHM myMap(1024);
  EXPECT_TRUE(myMap.begin() == myMap.end());

  for (int i = 0; i < 100; ++i) {
    myMap.insert(make_pair(i, folly::to<string>(i)));
  }
  for (int i = 0; i < 100; ++i) {
    EXPECT_EQ(myMap.find(i)->second, folly::to<string>(i));
  }

  myMap.insert(std::make_pair(999, "A"));
  myMap.insert(std::make_pair(999, "B"));
  EXPECT_EQ(myMap.find(999)->second, "A"); // shouldn't have overwritten
  myMap.find(999)->second = "B";
  myMap.find(999)->second = "C";
  EXPECT_EQ(myMap.find(999)->second, "C");
  EXPECT_EQ(myMap.find(999)->first, 999);
}

TEST(Ahm, BasicNoncopyable) {
  typedef AtomicHashMap<int64_t, std::unique_ptr<int>> AHM;
  AHM myMap(1024);
  EXPECT_TRUE(myMap.begin() == myMap.end());

  for (int i = 0; i < 50; ++i) {
    myMap.insert(make_pair(i, std::make_unique<int>(i)));
  }
  for (int i = 50; i < 100; ++i) {
    myMap.insert(i, std::make_unique<int>(i));
  }
  for (int i = 100; i < 150; ++i) {
    myMap.emplace(i, new int(i));
  }
  for (int i = 150; i < 200; ++i) {
    myMap.emplace(i, new int(i), std::default_delete<int>());
  }
  for (int i = 0; i < 200; ++i) {
    EXPECT_EQ(*(myMap.find(i)->second), i);
  }
  for (int i = 0; i < 200; i += 4) {
    myMap.erase(i);
  }
  for (int i = 0; i < 200; i += 4) {
    EXPECT_EQ(myMap.find(i), myMap.end());
  }
}

typedef int32_t KeyT;
typedef int32_t ValueT;

typedef AtomicHashMap<KeyT, ValueT> AHMapT;
typedef AHMapT::value_type RecordT;
typedef AtomicHashArray<KeyT, ValueT> AHArrayT;
AHArrayT::Config config;
typedef folly::QuadraticProbingAtomicHashMap<KeyT, ValueT> QPAHMapT;
QPAHMapT::Config qpConfig;
static AHArrayT::SmartPtr globalAHA(nullptr);
static std::unique_ptr<AHMapT> globalAHM;
static std::unique_ptr<QPAHMapT> globalQPAHM;

// Generate a deterministic value based on an input key
static int genVal(int key) {
  return key / 3;
}

static bool legalKey(const char* a);

struct EqTraits {
  bool operator()(const char* a, const char* b) {
    return legalKey(a) && (strcmp(a, b) == 0);
  }
  bool operator()(const char* a, const char& b) {
    return legalKey(a) && (a[0] != '\0') && (a[0] == b);
  }
  bool operator()(const char* a, const StringPiece b) {
    return legalKey(a) && (strlen(a) == b.size()) &&
        (strcmp(a, b.begin()) == 0);
  }
};

struct HashTraits {
  size_t operator()(const char* a) {
    size_t result = 0;
    while (a[0] != 0) {
      result += static_cast<size_t>(*(a++));
    }
    return result;
  }
  size_t operator()(const char& a) {
    return static_cast<size_t>(a);
  }
  size_t operator()(const StringPiece a) {
    size_t result = 0;
    for (const auto& ch : a) {
      result += static_cast<size_t>(ch);
    }
    return result;
  }
};

typedef AtomicHashMap<const char*, int64_t, HashTraits, EqTraits> AHMCstrInt;
AHMCstrInt::Config cstrIntCfg;

static bool legalKey(const char* a) {
  return a != cstrIntCfg.emptyKey && a != cstrIntCfg.lockedKey &&
      a != cstrIntCfg.erasedKey;
}

TEST(Ahm, BasicLookup) {
  AHMCstrInt myMap(1024, cstrIntCfg);
  EXPECT_TRUE(myMap.begin() == myMap.end());
  myMap.insert(std::make_pair("f", 42));
  EXPECT_EQ(42, myMap.find("f")->second);
  {
    // Look up a single char, successfully.
    auto it = myMap.find<char>('f');
    EXPECT_EQ(42, it->second);
  }
  {
    // Look up a single char, unsuccessfully.
    auto it = myMap.find<char>('g');
    EXPECT_TRUE(it == myMap.end());
  }
  {
    // Look up a string piece, successfully.
    const StringPiece piece("f");
    auto it = myMap.find(piece);
    EXPECT_EQ(42, it->second);
  }
}

TEST(Ahm, grow) {
  VLOG(1) << "Overhead: " << sizeof(AHArrayT) << " (array) "
          << sizeof(AHMapT) + sizeof(AHArrayT) << " (map/set) Bytes.";
  uint64_t numEntries = 10000;
  float sizeFactor = 0.46f;

  std::unique_ptr<AHMapT> m(new AHMapT(int(numEntries * sizeFactor), config));

  // load map - make sure we succeed and the index is accurate
  bool success = true;
  for (uint64_t i = 0; i < numEntries; i++) {
    auto ret = m->insert(RecordT(i, genVal(i)));
    success &= ret.second;
    success &= (m->findAt(ret.first.getIndex())->second == genVal(i));
  }
  // Overwrite vals to make sure there are no dups
  // Every insert should fail because the keys are already in the map.
  success = true;
  for (uint64_t i = 0; i < numEntries; i++) {
    auto ret = m->insert(RecordT(i, genVal(i * 2)));
    success &= (ret.second == false); // fail on collision
    success &= (ret.first->second == genVal(i)); // return the previous value
    success &= (m->findAt(ret.first.getIndex())->second == genVal(i));
  }
  EXPECT_TRUE(success);

  // check correctness
  EXPECT_GT(m->numSubMaps(), 1); // make sure we grew
  success = true;
  EXPECT_EQ(m->size(), numEntries);
  for (size_t i = 0; i < numEntries; i++) {
    success &= (m->find(i)->second == genVal(i));
  }
  EXPECT_TRUE(success);

  // Check findAt
  success = true;
  AHMapT::const_iterator retIt;
  for (int32_t i = 0; i < int32_t(numEntries); i++) {
    retIt = m->find(i);
    retIt = m->findAt(retIt.getIndex());
    success &= (retIt->second == genVal(i));
    // We use a uint32_t index so that this comparison is between two
    // variables of the same type.
    success &= (retIt->first == i);
  }
  EXPECT_TRUE(success);

  // Try modifying value
  m->find(8)->second = 5309;
  EXPECT_EQ(m->find(8)->second, 5309);

  // check clear()
  m->clear();
  success = true;
  for (uint64_t i = 0; i < numEntries / 2; i++) {
    success &= m->insert(RecordT(i, genVal(i))).second;
  }
  EXPECT_TRUE(success);
  EXPECT_EQ(m->size(), numEntries / 2);
}

TEST(Ahm, iterator) {
  int numEntries = 10000;
  float sizeFactor = .46f;
  std::unique_ptr<AHMapT> m(new AHMapT(int(numEntries * sizeFactor), config));

  // load map - make sure we succeed and the index is accurate
  for (int i = 0; i < numEntries; i++) {
    m->insert(RecordT(i, genVal(i)));
  }

  bool success = true;
  int count = 0;
  FOR_EACH (it, *m) {
    success &= (it->second == genVal(it->first));
    ++count;
  }
  EXPECT_TRUE(success);
  EXPECT_EQ(count, numEntries);
}

class Counters {
 private:
  // Note: Unfortunately can't currently put a std::atomic<int64_t> in
  // the value in ahm since it doesn't support types that are both non-copy
  // and non-move constructible yet.
  AtomicHashMap<int64_t, int64_t> ahm;

 public:
  explicit Counters(size_t numCounters) : ahm(numCounters) {}

  void increment(int64_t obj_id) {
    auto ret = ahm.insert(std::make_pair(obj_id, 1));
    if (!ret.second) {
      // obj_id already exists, increment count
      __sync_fetch_and_add(&ret.first->second, 1);
    }
  }

  int64_t getValue(int64_t obj_id) {
    auto ret = ahm.find(obj_id);
    return ret != ahm.end() ? ret->second : 0;
  }

  // export the counters without blocking increments
  string toString() {
    string ret = "{\n";
    ret.reserve(ahm.size() * 32);
    for (const auto& e : ahm) {
      ret += folly::to<string>("  [", e.first, ":", e.second, "]\n");
    }
    ret += "}\n";
    return ret;
  }
};

// If you get an error "terminate called without an active exception", there
// might be too many threads getting created - decrease numKeys and/or mult.
TEST(Ahm, counter) {
  const int numKeys = 10;
  const int mult = 10;
  Counters c(numKeys);
  vector<int64_t> keys;
  FOR_EACH_RANGE (i, 1, numKeys) { keys.push_back(i); }
  vector<std::thread> threads;
  for (auto key : keys) {
    FOR_EACH_RANGE (i, 0, key * mult) {
      threads.push_back(std::thread([&, key] { c.increment(key); }));
    }
  }
  for (auto& t : threads) {
    t.join();
  }
  string str = c.toString();
  for (auto key : keys) {
    int val = key * mult;
    EXPECT_EQ(val, c.getValue(key));
    EXPECT_NE(
        string::npos, str.find(folly::to<string>("[", key, ":", val, "]")));
  }
}

class Integer {
 public:
  explicit Integer(KeyT v = 0) : v_(v) {}

  Integer& operator=(const Integer& a) {
    static bool throwException_ = false;
    throwException_ = !throwException_;
    if (throwException_) {
      throw 1;
    }
    v_ = a.v_;
    return *this;
  }

  bool operator==(const Integer& a) const {
    return v_ == a.v_;
  }

 private:
  KeyT v_;
};

TEST(Ahm, map_exception_safety) {
  typedef AtomicHashMap<KeyT, Integer> MyMapT;

  int numEntries = 10000;
  float sizeFactor = 0.46f;
  std::unique_ptr<MyMapT> m(new MyMapT(int(numEntries * sizeFactor)));

  bool success = true;
  int count = 0;
  for (int i = 0; i < numEntries; i++) {
    try {
      m->insert(i, Integer(genVal(i)));
      success &= (m->find(i)->second == Integer(genVal(i)));
      ++count;
    } catch (...) {
      success &= !m->count(i);
    }
  }
  EXPECT_EQ(count, m->size());
  EXPECT_TRUE(success);
}

TEST(Ahm, basicErase) {
  size_t numEntries = 3000;

  std::unique_ptr<AHMapT> s(new AHMapT(numEntries, config));
  // Iterate filling up the map and deleting all keys a few times
  // to test more than one subMap.
  for (int iterations = 0; iterations < 4; ++iterations) {
    // Testing insertion of keys
    bool success = true;
    for (size_t i = 0; i < numEntries; ++i) {
      success &= !(s->count(i));
      auto ret = s->insert(RecordT(i, i));
      success &= s->count(i);
      success &= ret.second;
    }
    EXPECT_TRUE(success);
    EXPECT_EQ(s->size(), numEntries);

    // Delete every key in the map and verify that the key is gone and the the
    // size is correct.
    success = true;
    for (size_t i = 0; i < numEntries; ++i) {
      success &= s->erase(i);
      success &= (s->size() == numEntries - 1 - i);
      success &= !(s->count(i));
      success &= !(s->erase(i));
    }
    EXPECT_TRUE(success);
  }
  VLOG(1) << "Final number of subMaps = " << s->numSubMaps();
}

namespace {

inline KeyT randomizeKey(int key) {
  // We deterministically randomize the key to more accurately simulate
  // real-world usage, and to avoid pathalogical performance patterns (e.g.
  // those related to std::hash<int64_t>()(1) == 1).
  //
  // Use a hash function we don't normally use for ints to avoid interactions.
  return folly::hash::jenkins_rev_mix32(key);
}

int numOpsPerThread = 0;

void* insertThread(void* jj) {
  int64_t j = (int64_t)jj;
  for (int i = 0; i < numOpsPerThread; ++i) {
    KeyT key = randomizeKey(i + j * numOpsPerThread);
    globalAHM->insert(key, genVal(key));
  }
  return nullptr;
}

void* qpInsertThread(void* jj) {
  int64_t j = (int64_t)jj;
  for (int i = 0; i < numOpsPerThread; ++i) {
    KeyT key = randomizeKey(i + j * numOpsPerThread);
    globalQPAHM->insert(key, genVal(key));
  }
  return nullptr;
}

void* insertThreadArr(void* jj) {
  int64_t j = (int64_t)jj;
  for (int i = 0; i < numOpsPerThread; ++i) {
    KeyT key = randomizeKey(i + j * numOpsPerThread);
    globalAHA->insert(std::make_pair(key, genVal(key)));
  }
  return nullptr;
}

std::atomic<bool> runThreadsCreatedAllThreads;
void runThreads(void* (*mainFunc)(void*), int numThreads, void** statuses) {
  folly::BenchmarkSuspender susp;
  runThreadsCreatedAllThreads.store(false);
  vector<std::thread> threads;
  for (int64_t j = 0; j < numThreads; j++) {
    threads.emplace_back([statuses, mainFunc, j]() {
      auto ret = mainFunc((void*)j);
      if (statuses != nullptr) {
        statuses[j] = ret;
      }
    });
  }
  susp.dismiss();

  runThreadsCreatedAllThreads.store(true);
  for (size_t i = 0; i < threads.size(); ++i) {
    threads[i].join();
  }
}

void runThreads(void* (*mainFunc)(void*)) {
  runThreads(mainFunc, FLAGS_numThreads, nullptr);
}

} // namespace

TEST(Ahm, collision_test) {
  const int numInserts = 1000000 / 4;

  // Doing the same number on each thread so we collide.
  numOpsPerThread = numInserts;

  float sizeFactor = 0.46f;
  int entrySize = sizeof(KeyT) + sizeof(ValueT);
  VLOG(1) << "Testing " << numInserts << " unique " << entrySize
          << " Byte entries replicated in " << FLAGS_numThreads
          << " threads with " << FLAGS_maxLoadFactor * 100.0
          << "% max load factor.";

  globalAHM = std::make_unique<AHMapT>(int(numInserts * sizeFactor), config);

  size_t sizeInit = globalAHM->capacity();
  VLOG(1) << "  Initial capacity: " << sizeInit;

  double start = nowInUsec();
  runThreads([](void*) -> void* { // collisionInsertThread
    for (int i = 0; i < numOpsPerThread; ++i) {
      KeyT key = randomizeKey(i);
      globalAHM->insert(key, genVal(key));
    }
    return nullptr;
  });
  double elapsed = nowInUsec() - start;

  size_t finalCap = globalAHM->capacity();
  size_t sizeAHM = globalAHM->size();
  VLOG(1) << elapsed / sizeAHM << " usec per " << FLAGS_numThreads
          << " duplicate inserts (atomic).";
  VLOG(1) << "  Final capacity: " << finalCap << " in "
          << globalAHM->numSubMaps() << " sub maps ("
          << sizeAHM * 100 / finalCap << "% load factor, "
          << (finalCap - sizeInit) * 100 / sizeInit << "% growth).";

  // check correctness
  EXPECT_EQ(sizeAHM, numInserts);
  bool success = true;
  for (int i = 0; i < numInserts; ++i) {
    KeyT key = randomizeKey(i);
    success &= (globalAHM->find(key)->second == genVal(key));
  }
  EXPECT_TRUE(success);

  // check colliding finds
  start = nowInUsec();
  runThreads([](void*) -> void* { // collisionFindThread
    KeyT key(0);
    for (int i = 0; i < numOpsPerThread; ++i) {
      globalAHM->find(key);
    }
    return nullptr;
  });

  elapsed = nowInUsec() - start;

  VLOG(1) << elapsed / sizeAHM << " usec per " << FLAGS_numThreads
          << " duplicate finds (atomic).";
}

namespace {

const int kInsertPerThread = 100000;
int raceFinalSizeEstimate;

void* raceIterateThread(void*) {
  int count = 0;

  AHMapT::iterator it = globalAHM->begin();
  AHMapT::iterator end = globalAHM->end();
  for (; it != end; ++it) {
    ++count;
    if (count > raceFinalSizeEstimate) {
      EXPECT_FALSE("Infinite loop in iterator.");
      return nullptr;
    }
  }
  return nullptr;
}

void* raceInsertRandomThread(void*) {
  for (int i = 0; i < kInsertPerThread; ++i) {
    KeyT key = rand();
    globalAHM->insert(key, genVal(key));
  }
  return nullptr;
}

} // namespace

// Test for race conditions when inserting and iterating at the same time and
// creating multiple submaps.
TEST(Ahm, race_insert_iterate_thread_test) {
  const int kInsertThreads = 20;
  const int kIterateThreads = 20;
  raceFinalSizeEstimate = kInsertThreads * kInsertPerThread;

  VLOG(1) << "Testing iteration and insertion with " << kInsertThreads
          << " threads inserting and " << kIterateThreads
          << " threads iterating.";

  globalAHM = std::make_unique<AHMapT>(raceFinalSizeEstimate / 9, config);

  vector<pthread_t> threadIds;
  for (int j = 0; j < kInsertThreads + kIterateThreads; j++) {
    pthread_t tid;
    void* (*thread)(void*) =
        (j < kInsertThreads ? raceInsertRandomThread : raceIterateThread);
    if (pthread_create(&tid, nullptr, thread, nullptr) != 0) {
      LOG(ERROR) << "Could not start thread";
    } else {
      threadIds.push_back(tid);
    }
  }
  for (size_t i = 0; i < threadIds.size(); ++i) {
    pthread_join(threadIds[i], nullptr);
  }
  VLOG(1) << "Ended up with " << globalAHM->numSubMaps() << " submaps";
  VLOG(1) << "Final size of map " << globalAHM->size();
}

namespace {

const int kTestEraseInsertions = 200000;
std::atomic<int32_t> insertedLevel;

void* testEraseInsertThread(void*) {
  for (int i = 0; i < kTestEraseInsertions; ++i) {
    KeyT key = randomizeKey(i);
    globalAHM->insert(key, genVal(key));
    insertedLevel.store(i, std::memory_order_release);
  }
  insertedLevel.store(kTestEraseInsertions, std::memory_order_release);
  return nullptr;
}

void* testEraseEraseThread(void*) {
  for (int i = 0; i < kTestEraseInsertions; ++i) {
    /*
     * Make sure that we don't get ahead of the insert thread, because
     * part of the condition for this unit test succeeding is that the
     * map ends up empty.
     *
     * Note, there is a subtle case here when a new submap is
     * allocated: the erasing thread might get 0 from count(key)
     * because it hasn't seen numSubMaps_ update yet.  To avoid this
     * race causing problems for the test (it's ok for real usage), we
     * lag behind the inserter by more than just element.
     */
    const int lag = 10;
    int currentLevel;
    do {
      currentLevel = insertedLevel.load(std::memory_order_acquire);
      if (currentLevel == kTestEraseInsertions) {
        currentLevel += lag + 1;
      }
    } while (currentLevel - lag < i);

    KeyT key = randomizeKey(i);
    while (globalAHM->count(key)) {
      if (globalAHM->erase(key)) {
        break;
      }
    }
  }
  return nullptr;
}

} // namespace

// Here we have a single thread inserting some values, and several threads
// racing to delete the values in the order they were inserted.
TEST(Ahm, thread_erase_insert_race) {
  const int kInsertThreads = 1;
  const int kEraseThreads = 10;

  VLOG(1) << "Testing insertion and erase with " << kInsertThreads
          << " thread inserting and " << kEraseThreads << " threads erasing.";

  globalAHM = std::make_unique<AHMapT>(kTestEraseInsertions / 4, config);

  vector<pthread_t> threadIds;
  for (int64_t j = 0; j < kInsertThreads + kEraseThreads; j++) {
    pthread_t tid;
    void* (*thread)(void*) =
        (j < kInsertThreads ? testEraseInsertThread : testEraseEraseThread);
    if (pthread_create(&tid, nullptr, thread, (void*)j) != 0) {
      LOG(ERROR) << "Could not start thread";
    } else {
      threadIds.push_back(tid);
    }
  }
  for (size_t i = 0; i < threadIds.size(); i++) {
    pthread_join(threadIds[i], nullptr);
  }

  EXPECT_TRUE(globalAHM->empty());
  EXPECT_EQ(globalAHM->size(), 0);

  VLOG(1) << "Ended up with " << globalAHM->numSubMaps() << " submaps";
}

// Repro for T#483734: Duplicate AHM inserts due to incorrect AHA return value.
typedef AtomicHashArray<int32_t, int32_t> AHA;
AHA::Config configRace;
auto atomicHashArrayInsertRaceArray = AHA::create(2, configRace);
void* atomicHashArrayInsertRaceThread(void* /* j */) {
  AHA* arr = atomicHashArrayInsertRaceArray.get();
  uintptr_t numInserted = 0;
  while (!runThreadsCreatedAllThreads.load()) {
    ;
  }
  for (int i = 0; i < 2; i++) {
    if (arr->insert(RecordT(randomizeKey(i), 0)).first != arr->end()) {
      numInserted++;
    }
  }
  return (void*)numInserted;
}
TEST(Ahm, atomic_hash_array_insert_race) {
  AHA* arr = atomicHashArrayInsertRaceArray.get();
  int numIterations = 5000;
  constexpr int numThreads = 4;
  void* statuses[numThreads];
  for (int i = 0; i < numIterations; i++) {
    arr->clear();
    runThreads(atomicHashArrayInsertRaceThread, numThreads, statuses);
    EXPECT_GE(arr->size(), 1);
    for (int j = 0; j < numThreads; j++) {
      EXPECT_EQ(arr->size(), uintptr_t(statuses[j]));
    }
  }
}

// Repro for T#5841499. Race between erase() and find() on the same key.
TEST(Ahm, erase_find_race) {
  const uint64_t limit = 10000;
  AtomicHashMap<uint64_t, uint64_t> map(limit + 10);
  std::atomic<uint64_t> key{1};

  // Invariant: all values are equal to their keys.
  // At any moment there is one or two consecutive keys in the map.

  std::thread write_thread([&]() {
    while (true) {
      uint64_t k = ++key;
      if (k > limit) {
        break;
      }
      map.insert(k + 1, k + 1);
      map.erase(k);
    }
  });

  std::thread read_thread([&]() {
    while (true) {
      uint64_t k = key.load();
      if (k > limit) {
        break;
      }

      auto it = map.find(k);
      if (it != map.end()) {
        ASSERT_EQ(k, it->second);
      }
    }
  });

  read_thread.join();
  write_thread.join();
}

// Erase right after insert race bug repro (t9130653)
TEST(Ahm, erase_after_insert_race) {
  const uint64_t limit = 10000;
  const size_t num_threads = 100;
  const size_t num_iters = 500;
  AtomicHashMap<uint64_t, uint64_t> map(limit + 10);

  std::atomic<bool> go{false};
  std::vector<std::thread> ts;
  for (size_t i = 0; i < num_threads; ++i) {
    ts.emplace_back([&]() {
      while (!go) {
        continue;
      }
      for (size_t n = 0; n < num_iters; ++n) {
        map.erase(1);
        map.insert(1, 1);
      }
    });
  }

  go = true;

  for (auto& t : ts) {
    t.join();
  }
}

// Repro for a bug when iterator didn't skip empty submaps.
TEST(Ahm, iterator_skips_empty_submaps) {
  AtomicHashMap<uint64_t, uint64_t>::Config conf;
  conf.growthFactor = 1;

  AtomicHashMap<uint64_t, uint64_t> map(1, conf);

  map.insert(1, 1);
  map.insert(2, 2);
  map.insert(3, 3);

  map.erase(2);

  auto it = map.find(1);

  ASSERT_NE(map.end(), it);
  ASSERT_EQ(1, it->first);
  ASSERT_EQ(1, it->second);

  ++it;

  ASSERT_NE(map.end(), it);
  ASSERT_EQ(3, it->first);
  ASSERT_EQ(3, it->second);

  ++it;
  ASSERT_EQ(map.end(), it);
}

namespace {

void loadGlobalAha() {
  std::cout << "loading global AHA with " << FLAGS_numThreads
            << " threads...\n";
  uint64_t start = nowInUsec();
  globalAHA = AHArrayT::create(maxBMElements, config);
  numOpsPerThread = FLAGS_numBMElements / FLAGS_numThreads;
  CHECK_EQ(0, FLAGS_numBMElements % FLAGS_numThreads)
      << "kNumThreads must evenly divide kNumInserts.";
  runThreads(insertThreadArr);
  uint64_t elapsed = nowInUsec() - start;
  std::cout << "  took " << elapsed / 1000 << " ms ("
            << (elapsed * 1000 / FLAGS_numBMElements) << " ns/insert).\n";
  EXPECT_EQ(globalAHA->size(), FLAGS_numBMElements);
}

void loadGlobalAhm() {
  std::cout << "loading global AHM with " << FLAGS_numThreads
            << " threads...\n";
  uint64_t start = nowInUsec();
  globalAHM = std::make_unique<AHMapT>(maxBMElements, config);
  numOpsPerThread = FLAGS_numBMElements / FLAGS_numThreads;
  runThreads(insertThread);
  uint64_t elapsed = nowInUsec() - start;
  std::cout << "  took " << elapsed / 1000 << " ms ("
            << (elapsed * 1000 / FLAGS_numBMElements) << " ns/insert).\n";
  EXPECT_EQ(globalAHM->size(), FLAGS_numBMElements);
}

void loadGlobalQPAhm() {
  std::cout << "loading global QPAHM with " << FLAGS_numThreads
            << " threads...\n";
  uint64_t start = nowInUsec();
  globalQPAHM = std::make_unique<QPAHMapT>(maxBMElements, qpConfig);
  numOpsPerThread = FLAGS_numBMElements / FLAGS_numThreads;
  runThreads(qpInsertThread);
  uint64_t elapsed = nowInUsec() - start;
  std::cout << "  took " << elapsed / 1000 << " ms ("
            << (elapsed * 1000 / FLAGS_numBMElements) << " ns/insert).\n";
  EXPECT_EQ(globalQPAHM->size(), FLAGS_numBMElements);
}

} // namespace

BENCHMARK(st_aha_find, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  for (size_t i = 0; i < iters; i++) {
    KeyT key = randomizeKey(i);
    folly::doNotOptimizeAway(globalAHA->find(key)->second);
  }
}

BENCHMARK(st_ahm_find, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  for (size_t i = 0; i < iters; i++) {
    KeyT key = randomizeKey(i);
    folly::doNotOptimizeAway(globalAHM->find(key)->second);
  }
}

BENCHMARK(st_qpahm_find, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  for (size_t i = 0; i < iters; i++) {
    KeyT key = randomizeKey(i);
    folly::doNotOptimizeAway(globalQPAHM->find(key)->second);
  }
}

BENCHMARK_DRAW_LINE();

BENCHMARK(mt_ahm_miss, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  numOpsPerThread = iters / FLAGS_numThreads;
  runThreads([](void* jj) -> void* {
    int64_t j = (int64_t)jj;
    while (!runThreadsCreatedAllThreads.load()) {
      ;
    }
    for (int i = 0; i < numOpsPerThread; ++i) {
      KeyT key = i + j * numOpsPerThread * 100;
      folly::doNotOptimizeAway(globalAHM->find(key) == globalAHM->end());
    }
    return nullptr;
  });
}

BENCHMARK(mt_qpahm_miss, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  numOpsPerThread = iters / FLAGS_numThreads;
  runThreads([](void* jj) -> void* {
    int64_t j = (int64_t)jj;
    while (!runThreadsCreatedAllThreads.load()) {
      ;
    }
    for (int i = 0; i < numOpsPerThread; ++i) {
      KeyT key = i + j * numOpsPerThread * 100;
      folly::doNotOptimizeAway(globalQPAHM->find(key) == globalQPAHM->end());
    }
    return nullptr;
  });
}

BENCHMARK(st_ahm_miss, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  for (size_t i = 0; i < iters; i++) {
    KeyT key = randomizeKey(i + iters * 100);
    folly::doNotOptimizeAway(globalAHM->find(key) == globalAHM->end());
  }
}

BENCHMARK(st_qpahm_miss, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  for (size_t i = 0; i < iters; i++) {
    KeyT key = randomizeKey(i + iters * 100);
    folly::doNotOptimizeAway(globalQPAHM->find(key) == globalQPAHM->end());
  }
}

BENCHMARK(mt_ahm_find_insert_mix, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  numOpsPerThread = iters / FLAGS_numThreads;
  runThreads([](void* jj) -> void* {
    int64_t j = (int64_t)jj;
    while (!runThreadsCreatedAllThreads.load()) {
      ;
    }
    for (int i = 0; i < numOpsPerThread; ++i) {
      if (i % 128) { // ~1% insert mix
        KeyT key = randomizeKey(i + j * numOpsPerThread);
        folly::doNotOptimizeAway(globalAHM->find(key)->second);
      } else {
        KeyT key = randomizeKey(i + j * numOpsPerThread * 100);
        globalAHM->insert(key, genVal(key));
      }
    }
    return nullptr;
  });
}

BENCHMARK(mt_qpahm_find_insert_mix, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  numOpsPerThread = iters / FLAGS_numThreads;
  runThreads([](void* jj) -> void* {
    int64_t j = (int64_t)jj;
    while (!runThreadsCreatedAllThreads.load()) {
      ;
    }
    for (int i = 0; i < numOpsPerThread; ++i) {
      if (i % 128) { // ~1% insert mix
        KeyT key = randomizeKey(i + j * numOpsPerThread);
        folly::doNotOptimizeAway(globalQPAHM->find(key)->second);
      } else {
        KeyT key = randomizeKey(i + j * numOpsPerThread * 100);
        globalQPAHM->insert(key, genVal(key));
      }
    }
    return nullptr;
  });
}

BENCHMARK(mt_aha_find, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  numOpsPerThread = iters / FLAGS_numThreads;
  runThreads([](void* jj) -> void* {
    int64_t j = (int64_t)jj;
    while (!runThreadsCreatedAllThreads.load()) {
      ;
    }
    for (int i = 0; i < numOpsPerThread; ++i) {
      KeyT key = randomizeKey(i + j * numOpsPerThread);
      folly::doNotOptimizeAway(globalAHA->find(key)->second);
    }
    return nullptr;
  });
}

BENCHMARK(mt_ahm_find, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  numOpsPerThread = iters / FLAGS_numThreads;
  runThreads([](void* jj) -> void* {
    int64_t j = (int64_t)jj;
    while (!runThreadsCreatedAllThreads.load()) {
      ;
    }
    for (int i = 0; i < numOpsPerThread; ++i) {
      KeyT key = randomizeKey(i + j * numOpsPerThread);
      folly::doNotOptimizeAway(globalAHM->find(key)->second);
    }
    return nullptr;
  });
}

BENCHMARK(mt_qpahm_find, iters) {
  CHECK_LE(iters, FLAGS_numBMElements);
  numOpsPerThread = iters / FLAGS_numThreads;
  runThreads([](void* jj) -> void* {
    int64_t j = (int64_t)jj;
    while (!runThreadsCreatedAllThreads.load()) {
      ;
    }
    for (int i = 0; i < numOpsPerThread; ++i) {
      KeyT key = randomizeKey(i + j * numOpsPerThread);
      folly::doNotOptimizeAway(globalQPAHM->find(key)->second);
    }
    return nullptr;
  });
}

KeyT k;
BENCHMARK(st_baseline_modulus_and_random, iters) {
  for (size_t i = 0; i < iters; ++i) {
    k = randomizeKey(i) % iters;
  }
}

// insertions go last because they reset the map

BENCHMARK(mt_ahm_insert, iters) {
  BENCHMARK_SUSPEND {
    globalAHM = std::make_unique<AHMapT>(int(iters * LF), config);
    numOpsPerThread = iters / FLAGS_numThreads;
  }
  runThreads(insertThread);
}

BENCHMARK(mt_qpahm_insert, iters) {
  BENCHMARK_SUSPEND {
    globalQPAHM = std::make_unique<QPAHMapT>(int(iters * LF), qpConfig);
    numOpsPerThread = iters / FLAGS_numThreads;
  }
  runThreads(qpInsertThread);
}

BENCHMARK(st_ahm_insert, iters) {
  folly::BenchmarkSuspender susp;
  std::unique_ptr<AHMapT> ahm(new AHMapT(int(iters * LF), config));
  susp.dismiss();

  for (size_t i = 0; i < iters; i++) {
    KeyT key = randomizeKey(i);
    ahm->insert(key, genVal(key));
  }
}

BENCHMARK(st_qpahm_insert, iters) {
  folly::BenchmarkSuspender susp;
  std::unique_ptr<QPAHMapT> ahm(new QPAHMapT(int(iters * LF), qpConfig));
  susp.dismiss();

  for (size_t i = 0; i < iters; i++) {
    KeyT key = randomizeKey(i);
    ahm->insert(key, genVal(key));
  }
}

void benchmarkSetup() {
  config.maxLoadFactor = FLAGS_maxLoadFactor;
  qpConfig.maxLoadFactor = FLAGS_maxLoadFactor;
  configRace.maxLoadFactor = 0.5;
  int numCores = sysconf(_SC_NPROCESSORS_ONLN);
  loadGlobalAha();
  loadGlobalAhm();
  loadGlobalQPAhm();
  string numIters =
      folly::to<string>(std::min(1000000, int(FLAGS_numBMElements)));

  gflags::SetCommandLineOptionWithMode(
      "bm_max_iters", numIters.c_str(), gflags::SET_FLAG_IF_DEFAULT);
  gflags::SetCommandLineOptionWithMode(
      "bm_min_iters", numIters.c_str(), gflags::SET_FLAG_IF_DEFAULT);
  string numCoresStr = folly::to<string>(numCores);
  gflags::SetCommandLineOptionWithMode(
      "numThreads", numCoresStr.c_str(), gflags::SET_FLAG_IF_DEFAULT);

  std::cout << "\nRunning AHM benchmarks on machine with " << numCores
            << " logical cores.\n"
               "  num elements per map: "
            << FLAGS_numBMElements << "\n"
            << "  num threads for mt tests: " << FLAGS_numThreads << "\n"
            << "  AHM load factor: " << FLAGS_targetLoadFactor << "\n\n";
}

int main(int argc, char** argv) {
  testing::InitGoogleTest(&argc, argv);
  gflags::ParseCommandLineFlags(&argc, &argv, true);
  auto ret = RUN_ALL_TESTS();
  if (!ret && FLAGS_benchmark) {
    benchmarkSetup();
    folly::runBenchmarks();
  }
  return ret;
}

/*
loading global AHA with 8 threads...
  took 487 ms (40 ns/insert).
loading global AHM with 8 threads...
  took 478 ms (39 ns/insert).
loading global QPAHM with 8 threads...
  took 478 ms (39 ns/insert).

Running AHM benchmarks on machine with 24 logical cores.
  num elements per map: 12000000
  num threads for mt tests: 24
  AHM load factor: 0.75

============================================================================
folly/test/AtomicHashMapTest.cpp                relative  time/iter  iters/s
============================================================================
st_aha_find                                                 92.63ns   10.80M
st_ahm_find                                                107.78ns    9.28M
st_qpahm_find                                               90.69ns   11.03M
----------------------------------------------------------------------------
mt_ahm_miss                                                  2.09ns  477.36M
mt_qpahm_miss                                                1.37ns  728.82M
st_ahm_miss                                                241.07ns    4.15M
st_qpahm_miss                                              223.17ns    4.48M
mt_ahm_find_insert_mix                                       8.05ns  124.24M
mt_qpahm_find_insert_mix                                     9.10ns  109.85M
mt_aha_find                                                  6.82ns  146.68M
mt_ahm_find                                                  7.95ns  125.77M
mt_qpahm_find                                                6.81ns  146.83M
st_baseline_modulus_and_random                               6.02ns  166.03M
mt_ahm_insert                                               14.29ns   69.97M
mt_qpahm_insert                                             11.68ns   85.61M
st_ahm_insert                                              125.39ns    7.98M
st_qpahm_insert                                            128.76ns    7.77M
============================================================================
*/