[Embedding] Refactor ShrinkPolicy class and shrink functions in Embed…

…ding. (#861) Signed-off-by: lixy9474 <lxy268263@alibaba-inc.com>
DeepRec-AI · May 18, 2023 · 0c3e0ad · 0c3e0ad
1 parent f178454
commit 0c3e0ad
Show file tree

Hide file tree

Showing 14 changed files with 176 additions and 169 deletions.
diff --git a/tensorflow/core/framework/embedding/dram_leveldb_storage.h b/tensorflow/core/framework/embedding/dram_leveldb_storage.h
@@ -156,15 +156,9 @@ class DramLevelDBStore : public MultiTierStorage<K, V> {
     return Status::OK();
   }
 
-  Status Shrink(int64 value_len) override {
-    dram_->Shrink(value_len);
-    leveldb_->Shrink(value_len);
-    return Status::OK();
-  }
-
-  Status Shrink(int64 global_step, int64 steps_to_live) override {
-    dram_->Shrink(global_step, steps_to_live);
-    leveldb_->Shrink(global_step, steps_to_live);
+  Status Shrink(const ShrinkArgs& shrink_args) override {
+    dram_->Shrink(shrink_args);
+    leveldb_->Shrink(shrink_args);
     return Status::OK();
   }
 

diff --git a/tensorflow/core/framework/embedding/dram_pmem_storage.h b/tensorflow/core/framework/embedding/dram_pmem_storage.h
@@ -154,15 +154,9 @@ class DramPmemStorage : public MultiTierStorage<K, V> {
     return Status::OK();
   }
 
-  Status Shrink(int64 value_len) override {
-    dram_->Shrink(value_len);
-    pmem_->Shrink(value_len);
-    return Status::OK();
-  }
-
-  Status Shrink(int64 global_step, int64 steps_to_live) override {
-    dram_->Shrink(global_step, steps_to_live);
-    pmem_->Shrink(global_step, steps_to_live);
+  Status Shrink(const ShrinkArgs& shrink_args) override {
+    dram_->Shrink(shrink_args);
+    pmem_->Shrink(shrink_args);
     return Status::OK();
   }
 

diff --git a/tensorflow/core/framework/embedding/dram_ssd_storage.h b/tensorflow/core/framework/embedding/dram_ssd_storage.h
@@ -147,15 +147,9 @@ class DramSsdHashStorage : public MultiTierStorage<K, V> {
     return Status::OK();
   }
 
-  Status Shrink(int64 value_len) override {
-    dram_->Shrink(value_len);
-    ssd_hash_->Shrink(value_len);
-    return Status::OK();
-  }
-
-  Status Shrink(int64 global_step, int64 steps_to_live) override {
-    dram_->Shrink(global_step, steps_to_live);
-    ssd_hash_->Shrink(global_step, steps_to_live);
+  Status Shrink(const ShrinkArgs& shrink_args) override {
+    dram_->Shrink(shrink_args);
+    ssd_hash_->Shrink(shrink_args);
     return Status::OK();
   }
 

diff --git a/tensorflow/core/framework/embedding/embedding_var.h b/tensorflow/core/framework/embedding/embedding_var.h
@@ -373,10 +373,6 @@ class EmbeddingVar : public ResourceBase {
     return emb_config_.steps_to_live;
   }
 
-  float GetL2WeightThreshold() {
-    return emb_config_.l2_weight_threshold;
-  }
-
   bool IsMultiLevel() {
     return storage_->IsMultiLevel();
   }
@@ -553,13 +549,10 @@ class EmbeddingVar : public ResourceBase {
     return storage_;
   }
 
-  Status Shrink() {
-    return storage_->Shrink(value_len_);
-  }
-
-  Status Shrink(int64 gs) {
-    if (emb_config_.steps_to_live > 0) {
-      return storage_->Shrink(gs, emb_config_.steps_to_live);
+  Status Shrink(embedding::ShrinkArgs& shrink_args) {
+    if (emb_config_.is_primary()) {
+      shrink_args.value_len = value_len_;
+      return storage_->Shrink(shrink_args);
     } else {
       return Status::OK();
     }

diff --git a/tensorflow/core/framework/embedding/globalstep_shrink_policy.h b/tensorflow/core/framework/embedding/globalstep_shrink_policy.h
@@ -26,37 +26,44 @@ namespace embedding {
 template<typename K, typename V>
 class GlobalStepShrinkPolicy : public ShrinkPolicy<K, V> {
  public:
-  GlobalStepShrinkPolicy(
-      KVInterface<K, V>* kv,
-      Allocator* alloc,
-      int slot_num)
-      : ShrinkPolicy<K, V>(kv, alloc, slot_num) {}
+  GlobalStepShrinkPolicy(int64 steps_to_live,
+                         Allocator* alloc,
+                         KVInterface<K, V>* kv)
+      : steps_to_live_(steps_to_live),
+        kv_(kv),
+        ShrinkPolicy<K, V>(alloc) {}
 
   TF_DISALLOW_COPY_AND_ASSIGN(GlobalStepShrinkPolicy);
 
-  void Shrink(int64 global_step, int64 steps_to_live) {
-    ShrinkPolicy<K, V>::ReleaseDeleteValues();
-    ShrinkPolicy<K, V>::GetSnapshot();
-    FilterToDelete(global_step, steps_to_live);
+  void Shrink(const ShrinkArgs& shrink_args) override {
+    ShrinkPolicy<K, V>::ReleaseValuePtrs();
+    std::vector<K> key_list;
+    std::vector<ValuePtr<V>*> value_list;
+    kv_->GetSnapshot(&key_list, &value_list);
+    FilterToDelete(shrink_args.global_step,
+        key_list, value_list);
   }
 
  private:
-  void FilterToDelete(int64 global_step, int64 steps_to_live) {
-    for (int64 i = 0; i < ShrinkPolicy<K, V>::key_list_.size(); ++i) {
-      int64 version = ShrinkPolicy<K, V>::value_list_[i]->GetStep();
+  void FilterToDelete(int64 global_step,
+                      const std::vector<K>& key_list,
+                      const std::vector<ValuePtr<V>*>& value_list) {
+    for (int64 i = 0; i < key_list.size(); ++i) {
+      int64 version = value_list[i]->GetStep();
       if (version == -1) {
-        ShrinkPolicy<K, V>::value_list_[i]->SetStep(global_step);
+        value_list[i]->SetStep(global_step);
       } else {
-        if (global_step - version > steps_to_live) {
-          ShrinkPolicy<K, V>::kv_->Remove(ShrinkPolicy<K, V>::key_list_[i]);
-          ShrinkPolicy<K, V>::to_delete_.emplace_back(
-              ShrinkPolicy<K, V>::value_list_[i]);
+        if (global_step - version > steps_to_live_) {
+          kv_->Remove(key_list[i]);
+          ShrinkPolicy<K, V>::EmplacePointer(value_list[i]);
         }
       }
     }
-    ShrinkPolicy<K, V>::key_list_.clear();
-    ShrinkPolicy<K, V>::value_list_.clear();
   }
+
+ private:
+  int64 steps_to_live_;
+  KVInterface<K, V>* kv_;
 };
 } // embedding
 } // tensorflow

diff --git a/tensorflow/core/framework/embedding/hbm_dram_ssd_storage.h b/tensorflow/core/framework/embedding/hbm_dram_ssd_storage.h
@@ -387,17 +387,10 @@ class HbmDramSsdStorage : public MultiTierStorage<K, V> {
     LOG(FATAL)<<"HbmDramSsdStorage dosen't support GetSnaoshot.";
   }
 
-  Status Shrink(int64 value_len) override {
-    hbm_->Shrink(value_len);
-    dram_->Shrink(value_len);
-    ssd_->Shrink(value_len);
-    return Status::OK();
-  }
-
-  Status Shrink(int64 global_step, int64 steps_to_live) override {
-    hbm_->Shrink(global_step, steps_to_live);
-    dram_->Shrink(global_step, steps_to_live);
-    ssd_->Shrink(global_step, steps_to_live);
+  Status Shrink(const ShrinkArgs& shrink_args) override {
+    hbm_->Shrink(shrink_args);
+    dram_->Shrink(shrink_args);
+    ssd_->Shrink(shrink_args);
     return Status::OK();
   }
 

diff --git a/tensorflow/core/framework/embedding/hbm_dram_storage.h b/tensorflow/core/framework/embedding/hbm_dram_storage.h
@@ -352,15 +352,9 @@ class HbmDramStorage : public MultiTierStorage<K, V> {
     return temp_hbm_key_list.size() + temp_dram_key_list.size();
   }
 
-  Status Shrink(int64 value_len) override {
-    hbm_->Shrink(value_len);
-    dram_->Shrink(value_len);
-    return Status::OK();
-  }
-
-  Status Shrink(int64 global_step, int64 steps_to_live) override {
-    hbm_->Shrink(global_step, steps_to_live);
-    dram_->Shrink(global_step, steps_to_live);
+  Status Shrink(const ShrinkArgs& shrink_args) override {
+    hbm_->Shrink(shrink_args);
+    dram_->Shrink(shrink_args);
     return Status::OK();
   }
 

diff --git a/tensorflow/core/framework/embedding/l2weight_shrink_policy.h b/tensorflow/core/framework/embedding/l2weight_shrink_policy.h
@@ -27,46 +27,51 @@ template<typename K, typename V>
 class L2WeightShrinkPolicy : public ShrinkPolicy<K, V> {
  public:
   L2WeightShrinkPolicy(float l2_weight_threshold,
-      int64 primary_index, int64 primary_offset,
-      KVInterface<K, V>* kv, Allocator* alloc,
-      int slot_num)
-      : l2_weight_threshold_(l2_weight_threshold),
-        primary_index_(primary_index), primary_offset_(primary_offset),
-        ShrinkPolicy<K, V>(kv, alloc, slot_num) {}
+                       int64 index,
+                       int64 offset,
+                       Allocator* alloc,
+                       KVInterface<K, V>* kv)
+      : index_(index),
+        offset_(offset),
+        kv_(kv),
+        l2_weight_threshold_(l2_weight_threshold),
+        ShrinkPolicy<K, V>(alloc) {}
 
   TF_DISALLOW_COPY_AND_ASSIGN(L2WeightShrinkPolicy);
 
-  void Shrink(int64 value_len) {
-    ShrinkPolicy<K, V>::ReleaseDeleteValues();
-    ShrinkPolicy<K, V>::GetSnapshot();
-    FilterToDelete(value_len);
+  void Shrink(const ShrinkArgs& shrink_args) override {
+    ShrinkPolicy<K, V>::ReleaseValuePtrs();
+    std::vector<K> key_list;
+    std::vector<ValuePtr<V>*> value_list;
+    kv_->GetSnapshot(&key_list, &value_list);
+    FilterToDelete(shrink_args.value_len,
+                   key_list, value_list);
   }
 
- private: 
-  void FilterToDelete(int64 value_len) {
-    for (int64 i = 0; i < ShrinkPolicy<K, V>::key_list_.size(); ++i) {
-      V* val = ShrinkPolicy<K, V>::value_list_[i]->GetValue(
-          primary_index_, primary_offset_);
+ private:
+  void FilterToDelete(int64 value_len,
+                      const std::vector<K>& key_list,
+                      const std::vector<ValuePtr<V>*>& value_list) {
+    for (int64 i = 0; i < key_list.size(); ++i) {
+      V* val = value_list[i]->GetValue(index_, offset_);
       if (val != nullptr) {
         V l2_weight = (V)0.0;
         for (int64 j = 0; j < value_len; j++) {
             l2_weight += val[j] * val[j];
         }
         l2_weight *= (V)0.5;
         if (l2_weight < (V)l2_weight_threshold_) {
-          ShrinkPolicy<K, V>::kv_->Remove(ShrinkPolicy<K, V>::key_list_[i]);
-          ShrinkPolicy<K, V>::to_delete_.emplace_back(
-              ShrinkPolicy<K, V>::value_list_[i]);
+          kv_->Remove(key_list[i]);
+          ShrinkPolicy<K, V>::EmplacePointer(value_list[i]);
         }
       }
     }
-    ShrinkPolicy<K, V>::key_list_.clear();
-    ShrinkPolicy<K, V>::value_list_.clear();
   }
 
  private:
-  int64 primary_index_; // Shrink only handle primary slot
-  int64 primary_offset_;
+  int64 index_;
+  int64 offset_;
+  KVInterface<K, V>* kv_;
   float l2_weight_threshold_;
 };
 } // embedding

diff --git a/tensorflow/core/framework/embedding/shrink_policy.h b/tensorflow/core/framework/embedding/shrink_policy.h
@@ -23,40 +23,55 @@ namespace tensorflow {
 template<typename V>
 class ValuePtr;
 
+class Allocator;
+
 namespace embedding {
+struct ShrinkArgs {
+  ShrinkArgs(): global_step(0), value_len(0) {}
+
+  ShrinkArgs(int64 global_step,
+             int64 value_len)
+      : global_step(global_step),
+        value_len(value_len) {}
+  int64 global_step;
+  int64 value_len;
+};
+
 template<typename K, typename V>
 class ShrinkPolicy {
  public:
-  ShrinkPolicy(KVInterface<K, V>* kv, Allocator* alloc, int slot_num)
-      : kv_(kv), alloc_(alloc),
-        slot_num_(slot_num), shrink_count_(0) {}
+  ShrinkPolicy(Allocator* alloc): alloc_(alloc) {}
+  virtual ~ShrinkPolicy() {}
 
   TF_DISALLOW_COPY_AND_ASSIGN(ShrinkPolicy);
 
-  inline Status GetSnapshot() {
-    shrink_count_ = (shrink_count_ + 1) % slot_num_;
-    return kv_->GetSnapshot(&key_list_, &value_list_);
+  virtual void Shrink(const ShrinkArgs& shrink_args) = 0;
+
+ protected:
+  void EmplacePointer(ValuePtr<V>* value_ptr) {
+    to_delete_.emplace_back(value_ptr);
   }
-
-  void ReleaseDeleteValues() {
-    if (shrink_count_ == 0) {
-      for (auto it : to_delete_) {
-        it->Destroy(alloc_);
-        delete it;
-      }
-      to_delete_.clear();
+
+  void ReleaseValuePtrs() {
+    for (auto it : to_delete_) {
+      it->Destroy(alloc_);
+      delete it;
     }
+    to_delete_.clear();
   }
-
- protected: 
-  std::vector<K> key_list_;
-  std::vector<ValuePtr<V>*> value_list_;
+ protected:
   std::vector<ValuePtr<V>*> to_delete_;
-
-  KVInterface<K, V>* kv_;
+ private:
   Allocator* alloc_;
-  int slot_num_;
-  int shrink_count_;
+};
+
+template<typename K, typename V>
+class NonShrinkPolicy: public ShrinkPolicy<K, V> {
+ public:
+  NonShrinkPolicy(): ShrinkPolicy<K, V>(nullptr) {}
+  TF_DISALLOW_COPY_AND_ASSIGN(NonShrinkPolicy);
+
+  void Shrink(const ShrinkArgs& shrink_args) {}
 };
 } // embedding
 } // tensorflow