Feat/autolrfu

tensorflow/core/framework/embedding/cache.h

@@ -5,6 +5,7 @@
 #include <unordered_map>
 #include <set>
 #include <list>
+#include <deque>
 #include <limits>
 #include "tensorflow/core/framework/tensor.h"
 #include "tensorflow/core/platform/types.h"
@@ -129,9 +130,9 @@ class LRUCache : public BatchCache<K> {
 };
 
 template <class K>
-class LFUCache : public BatchCache<K> {
+class OriginLFUCache : public BatchCache<K> {
  public:
-  LFUCache() {
+  OriginLFUCache() {
     min_freq = std::numeric_limits<size_t>::max();
     max_freq = 0;
     freq_table.emplace_back(std::pair<std::list<LFUNode>*, int64>(
@@ -272,6 +273,369 @@ class LFUCache : public BatchCache<K> {
   mutex mu_;
 };
 
+
+template <class K>
+class LFUNode {
+ public:
+  K key;
+  size_t freq;
+  LFUNode(K key, unsigned now) : key(key), freq(0) {}
+  size_t GetIndex() { return freq; }
+  size_t UpdateAndReturnIndex(unsigned now, bool lru_mode) { return ++freq; }
+};
+
+template <class K>
+class AgingNode {
+ public:
+  static const uint8_t INIT_CNT = 5;
+  static const uint8_t MIN_CNT = 1;
+  static const uint8_t MAX_CNT = 255;
+  static const unsigned INCR_FACTOR = 7;
+  static const unsigned DECR_FACTOR = 10;
+  // 32640 = 1 + 2 + ... + 255
+  static const unsigned UNIT_STEP = 32640 * INCR_FACTOR;
+  static const unsigned IGNORE_STEP = 0;
+
+ public:
+  K key;
+  uint8_t count;
+  uint8_t index;
+  unsigned last;
+  AgingNode(K key, unsigned now)
+      : key(key), count(INIT_CNT), index(INIT_CNT), last(now) {}
+  AgingNode(typename std::list<AgingNode>::iterator that)
+      : key(that->key),
+        count(that->count),
+        index(that->index),
+        last(that->last) {}
+  void DecrByProb(unsigned period) {
+    size_t ret1 = rand();
+    size_t ret2 = RAND_MAX;
+    ret1 *= UNIT_STEP * DECR_FACTOR;
+    ret2 *= (period - IGNORE_STEP);
+    if (count > 0 && ret1 < ret2) count--;
+  }
+  void DecrByPeriod(unsigned period) {
+    unsigned decrease_value = period / UNIT_STEP;
+    if (decrease_value + INIT_CNT > count)
+      count = INIT_CNT;
+    else
+      count -= decrease_value;
+  }
+  void Decrease(unsigned now) {
+    unsigned period = now >= last
+                          ? now - last
+                          : std::numeric_limits<size_t>::max() - last + now;
+    DecrByPeriod(period);
+  }
+  void IncrByProb() {
+    size_t ret = rand();
+    ret *= (count - INIT_CNT) * INCR_FACTOR;
+    if (count < 255 && ret < RAND_MAX) count++;
+  }
+  uint8_t UpdateAndReturnIndex(unsigned now, bool lru_mode) {
+    Decrease(now);
+    IncrByProb();
+    if (lru_mode)
+      index = MAX_CNT;
+    else
+      index = count;
+    return index;
+  }
+  size_t GetIndex() { return index; }
+};
+
+template <class K, class Node>
+class BaseLFUCache : public BatchCache<K> {
+ public:
+  using map_iter =
+      typename std::unordered_map<K,
+                                  typename std::list<Node>::iterator>::iterator;
+  BaseLFUCache() {
+    min_freq = std::numeric_limits<size_t>::max();
+    max_freq = 0;
+    freq_table.emplace_back(
+        std::pair<std::list<Node>*, int64>(new std::list<Node>, 0));
+    BatchCache<K>::num_hit = 0;
+    BatchCache<K>::num_miss = 0;
+  }
+
+  size_t size() {
+    mutex_lock l(mu_);
+    return key_table.size();
+  }
+
+  size_t get_evic_ids(K* evic_ids, size_t k_size) {
+    mutex_lock l(mu_);
+    size_t true_size = 0;
+    size_t st_freq = min_freq;
+    for (size_t i = 0; i < k_size && key_table.size() > 0; ++i) {
+      auto rm_it = freq_table[st_freq].first->back();
+      key_table.erase(rm_it.key);
+      evic_ids[i] = rm_it.key;
+      ++true_size;
+      freq_table[st_freq].first->pop_back();
+      freq_table[st_freq].second--;
+      if (freq_table[st_freq].second == 0) {
+        ++st_freq;
+        while (st_freq <= max_freq) {
+          if (freq_table[st_freq].second == 0) {
+            ++st_freq;
+          } else {
+            break;
+          }
+        }
+      }
+    }
+    return true_size;
+  }
+
+  void AddNode(K id, unsigned now) {
+    Node node(id, now);
+    size_t index = node.GetIndex();
+    freq_table[index].first->emplace_front(node);
+    freq_table[index].second++;
+    key_table[id] = freq_table[index].first->begin();
+    min_freq = std::min(min_freq, index);
+    max_freq = std::max(max_freq, index);
+  }
+
+  void UpdateNode(K id, map_iter it, unsigned now, bool lru_mode) {
+    Node node = (Node)(*(it->second));
+    size_t index = node.GetIndex();
+    freq_table[index].first->erase(it->second);
+    freq_table[index].second--;
+    if (freq_table[index].second == 0) {
+      if (min_freq == index) min_freq += 1;
+    }
+    index = node.UpdateAndReturnIndex(now, lru_mode);
+    if (index == freq_table.size()) {
+      freq_table.emplace_back(
+          std::pair<std::list<Node>*, int64>(new std::list<Node>, 0));
+    }
+    max_freq = std::max(max_freq, index);
+    min_freq = std::min(min_freq, index);
+    freq_table[index].first->emplace_front(node);
+    freq_table[index].second++;
+    key_table[id] = freq_table[index].first->begin();
+  }
+
+  void add_to_rank(const K* batch_ids, size_t batch_size) {
+    mutex_lock l(mu_);
+    for (size_t i = 0; i < batch_size; ++i) {
+      K id = batch_ids[i];
+      auto it = key_table.find(id);
+      if (it == key_table.end()) {
+        AddNode(id, 0);
+        BatchCache<K>::num_miss++;
+      } else {
+        UpdateNode(id, it, 0, false);
+        BatchCache<K>::num_hit++;
+      }
+    }
+  }
+
+  void add_to_rank(const K* batch_ids, size_t batch_size,
+                   const int64* batch_version, const int64* batch_freqs) {
+    // TODO: add to rank accroding to the version of ids
+    add_to_rank(batch_ids, batch_size);
+  }
+
+ protected:
+  size_t min_freq;
+  size_t max_freq;
+  std::vector<std::pair<std::list<Node>*, int64>> freq_table;
+  std::unordered_map<K, typename std::list<Node>::iterator> key_table;
+  mutex mu_;
+};
+
+template <class K>
+class LFUCache : public BaseLFUCache<K, LFUNode<K>> {
+ public:
+  LFUCache() : BaseLFUCache<K, LFUNode<K>>() {}
+};
+
+template <class K>
+class AgingLFUCache : public BaseLFUCache<K, AgingNode<K>> {
+ public:
+  AgingLFUCache() : BaseLFUCache<K, AgingNode<K>>() { 
+    global_step = 0;
+    for (size_t i = 0; i < AgingNode<K>::MAX_CNT; ++i) {
+      this->freq_table.emplace_back(std::pair<std::list<AgingNode<K>>*, int64>(
+          new std::list<AgingNode<K>>, 0));
+    }
+  }
+
+  void add_to_rank(const K* batch_ids, size_t batch_size) {
+    mutex_lock l(this->mu_);
+    for (size_t i = 0; i < batch_size; ++i) {
+      if (global_step == std::numeric_limits<size_t>::max()) global_step = 0;
+      global_step++;
+      K id = batch_ids[i];
+      auto it = this->key_table.find(id);
+      if (it == this->key_table.end()) {
+        this->AddNode(id, global_step);
+        BatchCache<K>::num_miss++;
+      } else {
+        this->UpdateNode(id, it, global_step, false);
+        BatchCache<K>::num_hit++;
+      }
+    }
+  }
+
+  void add_to_rank(const K* batch_ids, size_t batch_size,
+                   const int64* batch_version, const int64* batch_freqs) {
+    // TODO: add to rank accroding to the version of ids
+    add_to_rank(batch_ids, batch_size);
+  }
+
+ protected:
+  size_t global_step;
+};
+
+template <class K>
+class AutoLRFUCache : public AgingLFUCache<K> {
+ public:
+  AutoLRFUCache(int64 cache_capacity) : AgingLFUCache<K>() {
+    cache_capacity_ = cache_capacity;
+    state = F0;
+    lru_mode = false;
+    prev_hit_rate = -100;
+    counter_replacement = 0;
+    num_replacement = 1;
+    HitSpan = 1000;
+  }
+
+  void rebuild() {
+    if (lru_mode) return;
+    std::unordered_map<K, AgingNode<K>*> new_table;
+    for (auto it = this->key_table.begin(); it != this->key_table.end(); it++) {
+      AgingNode<K>* node = new AgingNode<K>(it->second);
+      node->UpdateAndReturnIndex(this->global_step, lru_mode);
+      new_table[node->GetIndex()] = node;
+    }
+    this->freq_table.clear();
+    this->key_table.clear();
+    for (auto it = new_table.begin(); it != new_table.end(); it++) {
+      AgingNode<K> *node = (AgingNode<K>*)(it->second);
+      this->freq_table[node->GetIndex()].first->emplace_front(*node);
+      this->freq_table[node->GetIndex()].second++;
+      this->key_table[node->key] =
+          this->freq_table[node->GetIndex()].first->begin();
+    }
+  }
+
+  int get_hit_rate_len_100000(int len) {
+    int total = hit_recent.size();
+    if (total <= 0) return 0.0;
+    int hit = 0;
+    for (auto it = hit_recent.begin(); it != hit_recent.end(); it++)
+      hit += (len-- > 0) && (*it);
+    if (hit == 0) return 0.0;
+    return int(hit * 100000 / total);
+  }
+
+  int get_hit_rate100000() {
+    int total = hit_recent.size();
+    if (total <= 0) return 0.0;
+    int hit = 0;
+    for (auto it = hit_recent.begin(); it != hit_recent.end(); it++) hit += *it;
+    if (hit == 0) return 0.0;
+    return int(hit * 100000 / total);
+  }
+
+  void mode_switch() {
+    if (lru_mode) {
+      lru_mode = false;
+      rebuild();  // 根据保存的频次信息，重新构建链表
+    } else {
+      lru_mode = true;
+    }
+  }
+
+  void auto_switch() {
+    if (state == S3 && counter_replacement < HitSpan)
+      return;
+    else if (counter_replacement <
+             num_replacement * (cache_capacity_ + HitSpan))  // TODO:
+      return;
+
+    counter_replacement = 0;
+    int curr_hit_rate = get_hit_rate100000();
+    if (state == F0) {
+      if (prev_hit_rate >= 0 && curr_hit_rate < prev_hit_rate * 0.85) {
+        // (prev_hit_rate - curr_hit_rate) / prev_hit_rate > 0.15
+        mode_switch();
+        state = S1;
+      }
+    } else if (state == S1) {
+      if (curr_hit_rate <= 1.15 * prev_hit_rate) {  //没有显著提高
+        mode_switch();
+        state = F0;
+      } else {
+        state = S2;
+      }
+    } else if (state == S2) {
+      mode_switch();
+      state = S3;
+    } else if (state == S3) {
+      curr_hit_rate = get_hit_rate_len_100000(HitSpan);
+      if (curr_hit_rate > prev_hit_rate) {
+        state = F0;
+        num_replacement = 1;
+      } else {
+        mode_switch();
+        state = S2;
+        num_replacement = num_replacement * 2;
+      }
+    }
+    if (state != S3) prev_hit_rate = curr_hit_rate;
+  }
+
+  void add_to_rank(const K* batch_ids, size_t batch_size) {
+    mutex_lock l(this->mu_);
+    for (size_t i = 0; i < batch_size; ++i) {
+      auto_switch();
+      if (this->global_step == std::numeric_limits<size_t>::max())
+        this->global_step = 0;
+      this->global_step++;
+      if (hit_recent.size() > 6000) {
+        hit_recent.pop_back();
+      }
+
+      K id = batch_ids[i];
+      auto it = this->key_table.find(id);
+      if (it == this->key_table.end()) {
+        hit_recent.push_front(true);
+        this->AddNode(id, this->global_step);
+        BatchCache<K>::num_miss++;
+      } else {
+        hit_recent.push_front(false);
+        counter_replacement++;
+        this->UpdateNode(id, it, this->global_step, lru_mode);
+        BatchCache<K>::num_hit++;
+      }
+    }
+  }
+
+  void add_to_rank(const K* batch_ids, size_t batch_size,
+                   const int64* batch_version, const int64* batch_freqs) {
+    // TODO: add to rank accroding to the version of ids
+    add_to_rank(batch_ids, batch_size);
+  }
+
+ private:
+  enum State { F0, S1, S2, S3 };
+
+  int64 cache_capacity_;
+  std::deque<bool> hit_recent;
+  unsigned HitSpan;  // 每span次查看
+  size_t counter_replacement;
+  unsigned num_replacement;
+  int prev_hit_rate;
+  bool lru_mode;
+  State state;
+};
 } // embedding
 } // tensorflow