refine gpu memory swap performance

mindspore/ccsrc/backend/optimizer/mem_reuse/mem_copy_manager.h

@@ -19,6 +19,7 @@
 
 #include <vector>
 #include <map>
+#include <set>
 #include <queue>
 #include <memory>
 #include <utility>
@@ -40,29 +41,58 @@ struct TensorInfo {
 struct KernelExecutionInfo {
   size_t topo_order_{0};
   float execution_perform_{0.0};
-  bool trigger_swap_{false};
+  bool trigger_swap_out_{false};
-  bool need_swap_{false};
+  bool trigger_swap_in_{false};
-  // output index to topo orders of node users
+  size_t swap_in_task_num_{0};
+  // Key: output index, value: topo orders of node users
   std::map<size_t, std::vector<size_t>> node_users_map_;
-  // kernel output idx to host addr
+  // Key: output idx, value: (host addr, dirty or not)
-  std::map<size_t, HostAddress> host_addrs_;
+  std::map<size_t, std::pair<HostAddress, bool>> host_addrs_;
 
-  KernelExecutionInfo() : KernelExecutionInfo(0, 0.0, false, false) {}
+  KernelExecutionInfo() {}
-  explicit KernelExecutionInfo(size_t topo_order)
+  explicit KernelExecutionInfo(size_t topo_order) : KernelExecutionInfo(topo_order, 0.0, false, false, 0) {}
-      : topo_order_(topo_order), execution_perform_(0.0), trigger_swap_(false), need_swap_(false) {}
+  KernelExecutionInfo(size_t topo_order, float execution_perform, bool trigger_swap_out, bool trigger_swap_in,
-  KernelExecutionInfo(size_t topo_order, float execution_perform, bool trigger_swap, bool need_swap)
+                      size_t swap_in_task_num)
       : topo_order_(topo_order),
         execution_perform_(execution_perform),
-        trigger_swap_(trigger_swap),
+        trigger_swap_out_(trigger_swap_out),
-        need_swap_(need_swap) {}
+        trigger_swap_in_(trigger_swap_in),
+        swap_in_task_num_(swap_in_task_num) {}
 };
 
-// trigger swap
 struct MemSwapInfo {
   SwapKind swap_kind_;
-  // kernel need to be swapped
+  // Topo order of kernel need be swapped
-  AnfNodePtr kernel_{nullptr};
+  size_t topo_order_;
   size_t output_idx_{0};
+  // Record the swapping out position of swapping in tensor
+  size_t swap_out_pos_;
+};
+
+struct SwapInfoComp {
+  bool operator()(const MemSwapInfo &a, const MemSwapInfo &b) {
+    int swap_kind_a = static_cast<int>(a.swap_kind_);
+    int swap_kind_b = static_cast<int>(b.swap_kind_);
+    if (swap_kind_a < swap_kind_b) {
+      return true;
+    } else if (swap_kind_a > swap_kind_b) {
+      return false;
+    }
+
+    if (a.swap_out_pos_ < b.swap_out_pos_) {
+      return true;
+    } else if (a.swap_out_pos_ > b.swap_out_pos_) {
+      return false;
+    }
+
+    if (a.topo_order_ < b.topo_order_) {
+      return true;
+    } else if (a.topo_order_ > b.topo_order_) {
+      return false;
+    }
+
+    return a.output_idx_ < b.output_idx_;
+  }
 };
 
 class MemCopyManager {
@@ -90,6 +120,7 @@ class MemCopyManager {
   virtual void ClearSwapQueue() {}
 };
 using MemCopyManagerPtr = std::shared_ptr<MemCopyManager>;
+using MemSwapInfoSet = std::set<MemSwapInfo, SwapInfoComp>;
 }  // namespace memswap
 }  // namespace device
 }  // namespace mindspore

mindspore/ccsrc/backend/optimizer/mem_reuse/mem_swap_manager.cc

@@ -22,22 +22,17 @@
 namespace mindspore {
 namespace device {
 namespace memswap {
-void MemSwapManager::Init(const mindspore::session::KernelGraph *kernel_graph) {
+bool MemSwapManager::Init(const mindspore::session::KernelGraph *kernel_graph, size_t swap_mem_size) {
   MS_EXCEPTION_IF_NULL(kernel_graph);
   graph_manager_ = kernel_graph->manager();
   MS_EXCEPTION_IF_NULL(graph_manager_);
-  auto &kernels = kernel_graph->execution_order();
+  execution_order_ = kernel_graph->execution_order();
-  for (const auto &kernel : kernels) {
-    if (AnfAlgo::IsRealCNodeKernel(kernel) && (!opt::IsNopNode(kernel))) {
-      execution_order_.push_back(kernel);
-    }
-  }
 
   size_t kernel_index = 0;
   for (const auto &kernel : execution_order_) {
-    // parse topo order of kernel
+    // Parse topo order of kernel
     (void)kernel_execution_info_.emplace(kernel.get(), kernel_index++);
-    // parse tensor info
+    // Parse tensor info
     auto kernel_mod = AnfAlgo::GetKernelMod(kernel);
     MS_EXCEPTION_IF_NULL(kernel_mod);
     auto output_sizes = kernel_mod->GetOutputSizeList();
@@ -48,7 +43,7 @@ void MemSwapManager::Init(const mindspore::session::KernelGraph *kernel_graph) {
     }
   }
 
-  // parse topo order of user kernel
+  // Parse topo order of user kernel
   SaveUserKernelTopoOrder();
 
   sort(ordered_tensors_.begin(), ordered_tensors_.end(),
@@ -61,17 +56,103 @@ void MemSwapManager::Init(const mindspore::session::KernelGraph *kernel_graph) {
       tensor_size_num_++;
     }
   }
-  tensor_size_threshold_ = ordered_tensors_.front().tensor_size_;
+  if (!InitSwapThreshold(0)) {
-  tensor_size_threshold_idx_ = 0;
+    return false;
-
+  }
-  distance_threshold_ = kernel_index / kDistanceInitFactor;
   mem_swap_initialized_ = true;
   MS_EXCEPTION_IF_NULL(mem_copy_manager_);
   mem_copy_manager_->Init();
+  return true;
+}
+
+bool MemSwapManager::InitSwapThreshold(size_t swap_mem_size) {
+  distance_threshold_ = execution_order_.size() / kDistanceInitFactor;
+  distance_decay_step_ = execution_order_.size() / kDistanceInitFactor / tensor_size_num_;
+  if (distance_decay_step_ <= 1) {
+    distance_decay_step_ = 1;
+  }
+  tensor_size_threshold_ = ordered_tensors_.front().tensor_size_;
+  tensor_size_threshold_idx_ = 0;
+
+  size_t accumulation = 0;
+  while (accumulation < swap_mem_size) {
+    accumulation = 0;
+    for (const auto &tensor_info : ordered_tensors_) {
+      size_t tensor_size = tensor_info.tensor_size_;
+      if (tensor_size < tensor_size_threshold_) {
+        break;
+      }
+      if (!CheckDistanceBetweenKernels(tensor_info)) {
+        continue;
+      }
+
+      accumulation += tensor_info.tensor_size_;
+      if (accumulation >= swap_mem_size) {
+        return true;
+      }
+    }
+    RetreatSwapThreshold();
+    if (tensor_size_threshold_idx_ == ordered_tensors_.size() - 1 && distance_threshold_ < kDistanceLowerBound) {
+      MS_LOG(ERROR) << "Init swap threshold info failed";
+      return false;
+    }
+  }
+  return true;
+}
+
+void MemSwapManager::RetreatSwapThreshold() {
+  if (distance_threshold_ >= kDistanceLowerBound) {
+    bool update_one_decay_step = (distance_threshold_ > distance_decay_step_) &&
+                                 (distance_threshold_ - distance_decay_step_ >= kDistanceLowerBound);
+    if (update_one_decay_step) {
+      distance_threshold_ -= distance_decay_step_;
+    } else if (distance_threshold_ >= kDistanceLowerBound) {
+      size_t new_distance_decay_step = (distance_threshold_ - kDistanceLowerBound) / 4;
+      if (new_distance_decay_step < 1) {
+        new_distance_decay_step = 1;
+      }
+      distance_threshold_ -= new_distance_decay_step;
+    }
+  }
+
+  while (tensor_size_threshold_idx_ < ordered_tensors_.size() - 1) {
+    ++tensor_size_threshold_idx_;
+    if (tensor_size_threshold_ > ordered_tensors_[tensor_size_threshold_idx_].tensor_size_) {
+      tensor_size_threshold_ = ordered_tensors_[tensor_size_threshold_idx_].tensor_size_;
+      break;
+    }
+  }
+}
+
+bool MemSwapManager::CheckDistanceBetweenKernels(const TensorInfo &tensor_info) const {
+  const AnfNodePtr &kernel = tensor_info.kernel_;
+  auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
+  auto &node_users_map = kernel_exec_info.node_users_map_;
+
+  auto iter = node_users_map.find(tensor_info.output_idx_);
+  if (iter == node_users_map.end()) {
+    return false;
+  }
+
+  auto &node_users = iter->second;
+  if (node_users.front() - kernel_exec_info.topo_order_ > distance_threshold_) {
+    return true;
+  }
+
+  for (size_t i = 1; i < node_users.size(); ++i) {
+    if (node_users[i] - node_users[i - 1] > distance_threshold_) {
+      return true;
+    }
+  }
+  return false;
 }
 
 bool MemSwapManager::IsCommunicationRelevantOp(const AnfNodePtr &kernel) const {
   MS_EXCEPTION_IF_NULL(kernel);
+  if (AnfAlgo::IsCommunicationOp(kernel)) {
+    return true;
+  }
+
   NodeUsersMap &user_map = graph_manager_->node_users();
   auto iter = user_map.find(kernel);
   bool adjacent_with_communication_op = false;
@@ -81,7 +162,7 @@ bool MemSwapManager::IsCommunicationRelevantOp(const AnfNodePtr &kernel) const {
       node_set.begin(), node_set.end(),
       [](const std::pair<AnfNodePtr, int> &node_pair) { return AnfAlgo::IsCommunicationOp(node_pair.first); });
   }
-  return (AnfAlgo::IsCommunicationOp(kernel)) || adjacent_with_communication_op;
+  return adjacent_with_communication_op;
 }
 
 void MemSwapManager::SaveUserKernelTopoOrder() {
@@ -95,7 +176,7 @@ void MemSwapManager::SaveUserKernelTopoOrder() {
     auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
     for (auto &node_pair : node_set) {
       auto user_kernel = node_pair.first;
-      if (!AnfAlgo::IsRealCNodeKernel(user_kernel) || opt::IsNopNode(user_kernel)) {
+      if (!AnfAlgo::IsRealCNodeKernel(user_kernel)) {
         continue;
       }
 
@@ -138,21 +219,18 @@ void MemSwapManager::AddSwapInfo() {
     if (!need_swap) {
       continue;
     }
-    AddKernelNeedSwap(kernel, true);
     HostAddress host_addr;
     host_addr.size = tensor_size;
     auto ret = AllocHostPinnedMem(tensor_size, reinterpret_cast<void **>(&host_addr.addr));
     if (!ret) {
       MS_LOG(EXCEPTION) << "Alloc host pinned memory[" << tensor_size << "] failed.";
     }
-    kernel_exec_info.host_addrs_[output_idx] = host_addr;
+    kernel_exec_info.host_addrs_[output_idx] = std::make_pair(host_addr, true);
-    MemSwapInfo mem_swap_out_info = {SwapKind::kDeviceToHost, kernel, output_idx};
+    MemSwapInfo mem_swap_out_info = {SwapKind::kDeviceToHost, kernel_exec_info.topo_order_, output_idx, 0};
     if (node_users.size() > 1) {
       AddKernelMemSwapInfo(execution_order_[node_users[0]], mem_swap_out_info);
-      AddKernelTriggerSwap(execution_order_[node_users[0]], true);
     } else {
       AddKernelMemSwapInfo(kernel, mem_swap_out_info);
-      AddKernelTriggerSwap(kernel, true);
     }
 
     size_t swap_in_order = node_users.size() == 1 ? node_users[0] - 1 : node_users[1] - 1;
@@ -160,9 +238,8 @@ void MemSwapManager::AddSwapInfo() {
       MS_LOG(EXCEPTION) << "Select swap in point failed for op[" << AnfAlgo::GetCNodeName(kernel) << "]";
     }
     auto swap_in_kernel = execution_order_[swap_in_order];
-    MemSwapInfo mem_swap_in_info = {SwapKind::kHostToDevice, kernel, output_idx};
+    MemSwapInfo mem_swap_in_info = {SwapKind::kHostToDevice, kernel_exec_info.topo_order_, output_idx, 0};
     AddKernelMemSwapInfo(swap_in_kernel, mem_swap_in_info);
-    AddKernelTriggerSwap(swap_in_kernel, true);
 
     host_addrs_list_.push_back(host_addr);
   }
@@ -189,7 +266,7 @@ DeviceAddressPtr MemSwapManager::UpdateSwapQueue(SwapKind swap_kind) const {
   }
 }
 
-// retreat to find a workable swap scheme
+// Retreat to find a workable swap scheme
 bool MemSwapManager::RetreatSwapInfo() {
   if (!trigger_swap_) {
     trigger_swap_ = true;
@@ -220,6 +297,114 @@ bool MemSwapManager::RetreatSwapInfo() {
   return true;
 }
 
+void MemSwapManager::AdjustSwapInPos(const AnfNodePtr &kernel, size_t index) {
+  if (kernel_first_move_cache_map_.find(kernel.get()) == kernel_first_move_cache_map_.end()) {
+    CacheCurSwapInfoSet(kernel);
+  }
+
+  auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
+  size_t kernel_pos = kernel_exec_info.topo_order_;
+  auto &mem_swap_info = mem_swap_info_cache_list_[index];
+
+  if (QueryFirstTimeMovePos(kernel, index)) {
+    best_and_cur_pos_cache_.first = BestSwapInPerformPos(kernel, mem_swap_info);
+    best_and_cur_pos_cache_.second = best_and_cur_pos_cache_.first;
+    size_t best_pos = best_and_cur_pos_cache_.first;
+    if (best_pos != kernel_pos) {
+      MoveSwapInfoPos(best_pos, kernel_pos, mem_swap_info);
+    }
+    AddFirstTimeMovePos(kernel, index, false);
+    return;
+  }
+
+  auto &cur_pos = best_and_cur_pos_cache_.second;
+  if (cur_pos < kernel_pos) {
+    MoveSwapInfoPos(cur_pos + 1, cur_pos, mem_swap_info);
+    cur_pos++;
+  }
+}
+
+void MemSwapManager::CacheCurSwapInfoSet(const AnfNodePtr &kernel) {
+  if (!kernel_first_move_cache_map_.empty()) {
+    kernel_first_move_cache_map_.clear();
+  }
+  if (!mem_swap_info_cache_list_.empty()) {
+    mem_swap_info_cache_list_.clear();
+  }
+
+  auto mem_swap_info_set = QueryKernelMemSwapInfo(kernel);
+  size_t swap_in_task_cnt = 0;
+  for (auto &mem_swap_info : mem_swap_info_set) {
+    if (mem_swap_info.swap_kind_ == SwapKind::kHostToDevice) {
+      (void)mem_swap_info_cache_list_.push_back(mem_swap_info);
+      kernel_first_move_cache_map_[kernel.get()].push_back(true);
+      swap_in_task_cnt++;
+    }
+  }
+  size_t swap_in_task_num = QueryKernelTriggerSwapInTaskNum(kernel);
+  if (swap_in_task_cnt != swap_in_task_num) {
+    MS_LOG(EXCEPTION) << "Swap_in_task_cnt :" << swap_in_task_cnt
+                      << "must equal Swap_in_task_num: " << swap_in_task_num;
+  }
+}
+
+void MemSwapManager::AddFirstTimeMovePos(const AnfNodePtr &kernel, size_t index, bool first_time) {
+  auto iter = kernel_first_move_cache_map_.find(kernel.get());
+  if (iter == kernel_first_move_cache_map_.end()) {
+    MS_LOG(EXCEPTION) << "Can not find first time move pos info of op[" << AnfAlgo::GetCNodeName(kernel) << "]";
+  }
+  auto &first_move_list = iter->second;
+  if (index >= first_move_list.size()) {
+    MS_LOG(EXCEPTION) << "Index [" << index << "] out of range";
+  }
+  first_move_list[index] = first_time;
+}
+
+bool MemSwapManager::QueryFirstTimeMovePos(const AnfNodePtr &kernel, size_t index) const {
+  auto iter = kernel_first_move_cache_map_.find(kernel.get());
+  if (iter == kernel_first_move_cache_map_.end()) {
+    MS_LOG(EXCEPTION) << "Can not find first time move pos info of op[" << AnfAlgo::GetCNodeName(kernel) << "]";
+  }
+  const auto &first_move_list = iter->second;
+  if (index >= first_move_list.size()) {
+    MS_LOG(EXCEPTION) << "Index [" << index << "] out of range";
+  }
+  return first_move_list[index];
+}
+
+size_t MemSwapManager::BestSwapInPerformPos(const AnfNodePtr &trigger_kernel, const MemSwapInfo &mem_swap_info) const {
+  auto need_swap_kernel = QueryKerneByTopoOrder(mem_swap_info.topo_order_);
+  const PerformPair &perform_pair = QueryKernelSwapPerform(need_swap_kernel, mem_swap_info.output_idx_);
+  float swap_in_cost_time = perform_pair.second;
+  size_t swap_out_pos = mem_swap_info.swap_out_pos_;
+  auto &kernel_exec_info = SearchKernelExecutionInfo(trigger_kernel);
+  size_t trigger_kernel_pos = kernel_exec_info.topo_order_;
+  float kernel_execution_time = 0;
+
+  size_t pos = trigger_kernel_pos;
+  for (; pos > swap_out_pos + 1; pos--) {
+    auto kernel = QueryKerneByTopoOrder(pos - 1);
+    if (QueryKernelTriggerSwapIn(kernel)) {
+      return pos;
+    }
+    kernel_execution_time += QueryKernelExecutionPerform(QueryKerneByTopoOrder(pos));
+    if (kernel_execution_time >= swap_in_cost_time) {
+      return pos - 1;
+    }
+  }
+  return pos;
+}
+
+void MemSwapManager::MoveSwapInfoPos(size_t des_pos, size_t src_pos, const MemSwapInfo &mem_swap_info) {
+  if (des_pos == src_pos) {
+    MS_LOG(EXCEPTION) << "destination pos can not equal source pos";
+  }
+  auto des_kernel = QueryKerneByTopoOrder(des_pos);
+  auto src_kernel = QueryKerneByTopoOrder(src_pos);
+  AddKernelMemSwapInfo(des_kernel, mem_swap_info);
+  RemoveKernelMemSwapInfo(src_kernel, mem_swap_info);
+}
+
 KernelExecutionInfo &MemSwapManager::SearchKernelExecutionInfo(const AnfNodePtr &kernel) const {
   MS_EXCEPTION_IF_NULL(kernel);
   auto iter = kernel_execution_info_.find(kernel.get());
@@ -234,16 +419,6 @@ void MemSwapManager::AddKernelExecutionPerform(const AnfNodePtr &kernel, float p
   kernel_exec_info.execution_perform_ = perform;
 }
 
-void MemSwapManager::AddKernelTriggerSwap(const AnfNodePtr &kernel, bool trigger_swap) {
-  auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
-  kernel_exec_info.trigger_swap_ = trigger_swap;
-}
-
-void MemSwapManager::AddKernelNeedSwap(const AnfNodePtr &kernel, bool need_swap) {
-  auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
-  kernel_exec_info.need_swap_ = need_swap;
-}
-
 void MemSwapManager::AddKernelSwapPerform(const AnfNodePtr &kernel, size_t output_idx,
                                           const std::pair<float, float> &perform) {
   MS_EXCEPTION_IF_NULL(kernel);
@@ -252,7 +427,42 @@ void MemSwapManager::AddKernelSwapPerform(const AnfNodePtr &kernel, size_t outpu
 
 void MemSwapManager::AddKernelMemSwapInfo(const AnfNodePtr &kernel, const MemSwapInfo &mem_swap_info) {
   MS_EXCEPTION_IF_NULL(kernel);
-  mem_swap_info_[kernel.get()].push_back(mem_swap_info);
+  (void)mem_swap_info_map_[kernel.get()].insert(mem_swap_info);
+  auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
+  if (mem_swap_info.swap_kind_ == SwapKind::kDeviceToHost) {
+    kernel_exec_info.trigger_swap_out_ = true;
+  } else if (mem_swap_info.swap_kind_ == SwapKind::kHostToDevice) {
+    kernel_exec_info.swap_in_task_num_++;
+    kernel_exec_info.trigger_swap_in_ = true;
+  }
+}
+
+void MemSwapManager::RemoveKernelMemSwapInfo(const AnfNodePtr &kernel, const MemSwapInfo &mem_swap_info) {
+  MS_EXCEPTION_IF_NULL(kernel);
+  if (mem_swap_info.swap_kind_ == SwapKind::kHostToDevice) {
+    auto map_iter = mem_swap_info_map_.find(kernel.get());
+    if (map_iter == mem_swap_info_map_.end()) {
+      MS_LOG(EXCEPTION) << "Can not find memory swap information of op[" << AnfAlgo::GetCNodeName(kernel) << "]";
+    }
+    MemSwapInfoSet &mem_swap_info_set = map_iter->second;
+
+    auto set_iter = mem_swap_info_set.find(mem_swap_info);
+    if (set_iter == mem_swap_info_set.end()) {
+      MS_LOG(EXCEPTION) << "Can not find memory swap information in mem swap info set";
+    }
+    mem_swap_info_set.erase(set_iter);
+
+    auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
+    if (kernel_exec_info.swap_in_task_num_ > 0) {
+      kernel_exec_info.swap_in_task_num_--;
+    }
+    if (kernel_exec_info.swap_in_task_num_ == 0) {
+      kernel_exec_info.trigger_swap_in_ = false;
+    }
+    if (mem_swap_info_set.empty()) {
+      (void)mem_swap_info_map_.erase(kernel.get());
+    }
+  }
 }
 
 float MemSwapManager::QueryKernelExecutionPerform(const AnfNodePtr &kernel) const {
@@ -262,12 +472,24 @@ float MemSwapManager::QueryKernelExecutionPerform(const AnfNodePtr &kernel) cons
 
 bool MemSwapManager::QueryKernelTriggerSwap(const AnfNodePtr &kernel) const {
   const auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
-  return kernel_exec_info.trigger_swap_;
+  return kernel_exec_info.trigger_swap_out_ || kernel_exec_info.trigger_swap_in_;
 }
 
-bool MemSwapManager::QueryKernelNeedSwap(const AnfNodePtr &kernel) const {
+bool MemSwapManager::QueryKernelTriggerSwapIn(const AnfNodePtr &kernel) const {
   const auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
-  return kernel_exec_info.need_swap_;
+  return kernel_exec_info.trigger_swap_in_;
+}
+
+size_t MemSwapManager::QueryKernelTriggerSwapInTaskNum(const AnfNodePtr &kernel) const {
+  const auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
+  return kernel_exec_info.swap_in_task_num_;
+}
+
+const AnfNodePtr MemSwapManager::QueryKerneByTopoOrder(size_t index) const {
+  if (index >= execution_order_.size()) {
+    MS_LOG(EXCEPTION) << "Index [" << index << "] out of range";
+  }
+  return execution_order_[index];
 }
 
 const PerformPair &MemSwapManager::QueryKernelSwapPerform(const AnfNodePtr &kernel, size_t output_idx) const {
@@ -286,15 +508,70 @@ const PerformPair &MemSwapManager::QueryKernelSwapPerform(const AnfNodePtr &kern
   return iter_output->second;
 }
 
-const std::vector<MemSwapInfo> &MemSwapManager::QueryKernelMemSwapInfo(const AnfNodePtr &kernel) const {
+const MemSwapInfoSet &MemSwapManager::QueryKernelMemSwapInfo(const AnfNodePtr &kernel) const {
   MS_EXCEPTION_IF_NULL(kernel);
-  auto iter = mem_swap_info_.find(kernel.get());
+  auto iter = mem_swap_info_map_.find(kernel.get());
-  if (iter == mem_swap_info_.end()) {
+  if (iter == mem_swap_info_map_.end()) {
-    MS_LOG(EXCEPTION) << "Can not find memory swap information data of op[" << AnfAlgo::GetCNodeName(kernel) << "]";
+    MS_LOG(EXCEPTION) << "Can not find memory swap information of op[" << AnfAlgo::GetCNodeName(kernel) << "]";
   }
   return iter->second;
 }
 
+void MemSwapManager::AssignHostMemory() {
+  for (auto &kernel_exec_info_pair : kernel_execution_info_) {
+    auto &kernel_exec_info = kernel_exec_info_pair.second;
+    auto &host_addrs_map = kernel_exec_info.host_addrs_;
+    for (auto &host_addr_pair : host_addrs_map) {
+      auto &host_addr = host_addr_pair.second.first;
+      auto ret = AllocHostPinnedMem(host_addr.size, reinterpret_cast<void **>(&host_addr.addr));
+      if (!ret) {
+        MS_LOG(EXCEPTION) << "Alloc host pinned memory[" << host_addr.size << "] failed.";
+      }
+      host_addrs_list_.push_back(host_addr);
+    }
+  }
+}
+
+const HostAddress &MemSwapManager::QueryKernelHostAddr(const AnfNodePtr &kernel, size_t output_idx) const {
+  auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
+  auto &host_addrs = kernel_exec_info.host_addrs_;
+  auto iter = host_addrs.find(output_idx);
+  if (iter == host_addrs.end()) {
+    MS_LOG(EXCEPTION) << "Can not find host address of op[" << AnfAlgo::GetCNodeName(kernel) << "]";
+  }
+  return (iter->second).first;
+}
+
+void MemSwapManager::AddKernelHostAddrIsDirty(const AnfNodePtr &kernel, size_t output_idx, bool dirty) {
+  auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
+  auto &host_addrs = kernel_exec_info.host_addrs_;
+  auto iter = host_addrs.find(output_idx);
+  if (iter == host_addrs.end()) {
+    MS_LOG(EXCEPTION) << "Can not find host memory dirty info of op[" << AnfAlgo::GetCNodeName(kernel) << "]";
+  }
+  (iter->second).second = dirty;
+}
+
+bool MemSwapManager::QueryKernelHostAddrIsDirty(const AnfNodePtr &kernel, size_t output_idx) const {
+  auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
+  auto &host_addrs = kernel_exec_info.host_addrs_;
+  auto iter = host_addrs.find(output_idx);
+  if (iter == host_addrs.end()) {
+    MS_LOG(EXCEPTION) << "Can not find host memory dirty info of op[" << AnfAlgo::GetCNodeName(kernel) << "]";
+  }
+  return (iter->second).second;
+}
+
+void MemSwapManager::ResetHostAddrIsDirty() {
+  for (auto &kernel_exec_info_pair : kernel_execution_info_) {
+    auto &kernel_exec_info = kernel_exec_info_pair.second;
+    auto &host_addrs = kernel_exec_info.host_addrs_;
+    for (auto &host_addr : host_addrs) {
+      host_addr.second.second = true;
+    }
+  }
+}
+
 void MemSwapManager::InsertSwapInBlackList(const void *device_ptr) { swap_in_blacklist_.insert(device_ptr); }
 
 bool MemSwapManager::FindInSwapInBlackList(const void *device_ptr) const {
@@ -302,16 +579,6 @@ bool MemSwapManager::FindInSwapInBlackList(const void *device_ptr) const {
   return iter != swap_in_blacklist_.end();
 }
 
-const HostAddress &MemSwapManager::kernel_host_addr(const AnfNodePtr &kernel, size_t output_idx) const {
-  auto &kernel_exec_info = SearchKernelExecutionInfo(kernel);
-  auto &host_addrs = kernel_exec_info.host_addrs_;
-  auto iter = host_addrs.find(output_idx);
-  if (iter == host_addrs.end()) {
-    MS_LOG(EXCEPTION) << "Can not find host address of op[" << AnfAlgo::GetCNodeName(kernel) << "]";
-  }
-  return iter->second;
-}
-
 bool MemSwapManager::AllocHostPinnedMem(size_t size, void **addr) const {
   return mem_copy_manager_->AllocHostPinnedMem(size, addr);
 }
@@ -331,13 +598,14 @@ void MemSwapManager::ResetSwapInfo() {
   ClearSwapQueue();
   for (auto &kernel_exec_info_pair : kernel_execution_info_) {
     auto &kernel_exec_info = kernel_exec_info_pair.second;
-    kernel_exec_info.trigger_swap_ = false;
+    kernel_exec_info.trigger_swap_out_ = false;
-    kernel_exec_info.need_swap_ = false;
+    kernel_exec_info.trigger_swap_in_ = false;
+    kernel_exec_info.swap_in_task_num_ = 0;
     kernel_exec_info.host_addrs_.clear();
   }
   ReleaseHostPinnedMem();
   swap_in_blacklist_.clear();
-  mem_swap_info_.clear();
+  mem_swap_info_map_.clear();
 }
 }  // namespace memswap
 }  // namespace device

mindspore/ccsrc/backend/optimizer/mem_reuse/mem_swap_manager.h

@@ -32,7 +32,11 @@ namespace memswap {
 class MemSwapManager {
  public:
   explicit MemSwapManager(const MemCopyManagerPtr &mem_copy_manager)
-      : tensor_size_threshold_(0), tensor_size_threshold_idx_(0), tensor_size_num_(1), distance_threshold_(1) {
+      : tensor_size_threshold_(0),
+        tensor_size_threshold_idx_(0),
+        tensor_size_num_(1),
+        distance_threshold_(1),
+        distance_decay_step_(1) {
     mem_copy_manager_ = mem_copy_manager;
   }
 
@@ -42,7 +46,7 @@ class MemSwapManager {
 
   ~MemSwapManager() = default;
 
-  void Init(const mindspore::session::KernelGraph *kernel_graph);
+  bool Init(const mindspore::session::KernelGraph *kernel_graph, size_t swap_mem_size = 0);
 
   void AddMemSwapTask(SwapKind swap_kind, const DeviceAddressPtr &device_address,
                       const HostAddress &host_address) const;
@@ -51,9 +55,10 @@ class MemSwapManager {
 
   DeviceAddressPtr UpdateSwapQueue(SwapKind swap_kind) const;
 
-  // retreat to find a workable swap scheme
   bool RetreatSwapInfo();
 
+  void AdjustSwapInPos(const AnfNodePtr &kernel, size_t index);
+
   bool trigger_swap() const { return trigger_swap_; }
 
   bool mem_swap_init() const { return mem_swap_initialized_; }
@@ -70,16 +75,28 @@ class MemSwapManager {
 
   bool QueryKernelTriggerSwap(const AnfNodePtr &kernel) const;
 
-  bool QueryKernelNeedSwap(const AnfNodePtr &kernel) const;
+  bool QueryKernelTriggerSwapIn(const AnfNodePtr &kernel) const;
 
-  const std::vector<MemSwapInfo> &QueryKernelMemSwapInfo(const AnfNodePtr &kernel) const;
+  size_t QueryKernelTriggerSwapInTaskNum(const AnfNodePtr &kernel) const;
+
+  const AnfNodePtr QueryKerneByTopoOrder(size_t index) const;
+
+  const MemSwapInfoSet &QueryKernelMemSwapInfo(const AnfNodePtr &kernel) const;
+
+  void AssignHostMemory();
+
+  const HostAddress &QueryKernelHostAddr(const AnfNodePtr &kernel, size_t output_idx) const;
+
+  void AddKernelHostAddrIsDirty(const AnfNodePtr &kernel, size_t output_idx, bool dirty);
+
+  bool QueryKernelHostAddrIsDirty(const AnfNodePtr &kernel, size_t output_idx) const;
+
+  void ResetHostAddrIsDirty();
 
   void InsertSwapInBlackList(const void *device_ptr);
 
   bool FindInSwapInBlackList(const void *device_ptr) const;
 
-  const HostAddress &kernel_host_addr(const AnfNodePtr &kernel, size_t output_idx) const;
-
   bool AllocHostPinnedMem(size_t size, void **addr) const;
 
   void ReleaseHostPinnedMem();
@@ -93,27 +110,47 @@ class MemSwapManager {
 
   void SaveUserKernelTopoOrder();
 
-  void AddKernelTriggerSwap(const AnfNodePtr &kernel, bool trigger_swap);
+  bool InitSwapThreshold(size_t swap_mem_size);
 
-  void AddKernelNeedSwap(const AnfNodePtr &kernel, bool need_swap);
+  void RetreatSwapThreshold();
+
+  void CacheCurSwapInfoSet(const AnfNodePtr &kernel);
+
+  void AddFirstTimeMovePos(const AnfNodePtr &kernel, size_t index, bool first_time);
+
+  bool QueryFirstTimeMovePos(const AnfNodePtr &kernel, size_t index) const;
+
+  size_t BestSwapInPerformPos(const AnfNodePtr &trigger_kernel, const MemSwapInfo &mem_swap_info) const;
+
+  void MoveSwapInfoPos(size_t des_pos, size_t src_pos, const MemSwapInfo &mem_swap_info);
 
   void AddKernelMemSwapInfo(const AnfNodePtr &kernel, const MemSwapInfo &mem_swap_info);
 
+  void RemoveKernelMemSwapInfo(const AnfNodePtr &kernel, const MemSwapInfo &mem_swap_info);
+
+  bool CheckDistanceBetweenKernels(const TensorInfo &tensor_info) const;
+
   bool IsCommunicationRelevantOp(const AnfNodePtr &kernel) const;
 
   std::vector<CNodePtr> execution_order_;
   std::vector<TensorInfo> ordered_tensors_;
   std::unordered_map<void *, KernelExecutionInfo> kernel_execution_info_;
   std::unordered_map<void *, std::map<size_t, PerformPair>> kernel_swap_perform_;
-  // trigger swap kernel key : MemSwapInfo of kernel need to be swapped
+  // Key: trigger swap kernel, value: MemSwapInfoSet of kernel need to be swapped
-  std::unordered_map<void *, std::vector<MemSwapInfo>> mem_swap_info_;
+  std::unordered_map<void *, MemSwapInfoSet> mem_swap_info_map_;
   std::vector<HostAddress> host_addrs_list_;
   std::unordered_set<const void *> swap_in_blacklist_;
 
+  // Key: cache kernel address, value: lists of first time move pos or not
+  std::map<void *, std::vector<bool>> kernel_first_move_cache_map_;
+  std::vector<MemSwapInfo> mem_swap_info_cache_list_;
+  std::pair<size_t, size_t> best_and_cur_pos_cache_;
+
   size_t tensor_size_threshold_;
   size_t tensor_size_threshold_idx_;
   size_t tensor_size_num_;
   size_t distance_threshold_;
+  size_t distance_decay_step_;
 
   MemCopyManagerPtr mem_copy_manager_{nullptr};
   FuncGraphManagerPtr graph_manager_{nullptr};

mindspore/ccsrc/backend/session/anf_runtime_algorithm.cc

@@ -707,6 +707,18 @@ DeviceAddress *AnfRuntimeAlgorithm::GetWorkspaceAddr(const AnfNodePtr &node, siz
   return addr;
 }
 
+// get workspace device mutable addr of anf_node
+DeviceAddressPtr AnfRuntimeAlgorithm::GetMutableWorkspaceAddr(const AnfNodePtr &node, size_t index) {
+  MS_EXCEPTION_IF_NULL(node);
+  auto kernel_info = dynamic_cast<device::KernelInfo *>(node->kernel_info());
+  MS_EXCEPTION_IF_NULL(kernel_info);
+  auto addr = kernel_info->GetMutableWorkspaceAddr(index);
+  if (addr == nullptr) {
+    MS_LOG(EXCEPTION) << "Index " << index << " of node " << node->DebugString() << "] workspace addr is not exist";
+  }
+  return addr;
+}
+
 // set infer shapes and types of anf node
 void AnfRuntimeAlgorithm::SetOutputInferTypeAndShape(const std::vector<TypeId> &types,
                                                      const std::vector<std::vector<size_t>> &shapes, AnfNode *node) {

mindspore/ccsrc/backend/session/anf_runtime_algorithm.h

@@ -149,6 +149,8 @@ class AnfRuntimeAlgorithm {
   static void SetWorkspaceAddr(const DeviceAddressPtr &addr, size_t output_idx, AnfNode *node);
   // get workspace device addr of anf_node
   static DeviceAddress *GetWorkspaceAddr(const AnfNodePtr &node, size_t output_idx);
+  // get workspace device mutable addr of anf_node
+  static DeviceAddressPtr GetMutableWorkspaceAddr(const AnfNodePtr &node, size_t index);
   // set infer shapes and types of anf node
   static void SetOutputInferTypeAndShape(const std::vector<TypeId> &types,
                                          const std::vector<std::vector<size_t>> &shapes, AnfNode *node);

mindspore/ccsrc/runtime/device/gpu/cuda_driver.cc

@@ -209,6 +209,16 @@ bool CudaDriver::QueryEvent(const DeviceEvent &event) {
   }
 }
 
+bool CudaDriver::ElapsedTime(float *cost_time, const DeviceEvent &start, const DeviceEvent &end) {
+  auto ret = cudaEventElapsedTime(cost_time, (cudaEvent_t)start, (cudaEvent_t)end);
+  if (ret == cudaSuccess) {
+    return true;
+  } else {
+    MS_LOG(ERROR) << "cudaEventElapsedTime failed, ret[" << static_cast<int>(ret) << "], " << cudaGetErrorString(ret);
+    return false;
+  }
+}
+
 int CudaDriver::device_count() {
   int dev_count;
   auto ret = cudaGetDeviceCount(&dev_count);

mindspore/ccsrc/runtime/device/gpu/cuda_driver.h

@@ -57,6 +57,7 @@ class CudaDriver {
   static bool RecordEvent(DeviceEvent event, DeviceStream stream = 0);
   static bool SyncEvent(const DeviceEvent &event);
   static bool QueryEvent(const DeviceEvent &event);
+  static bool ElapsedTime(float *cost_time, const DeviceEvent &start, const DeviceEvent &end);
 
   // Encapsulate the cuda APIs associated with device management.
   static int device_count();

mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc

@@ -33,6 +33,7 @@
 namespace mindspore {
 namespace device {
 namespace gpu {
+using mindspore::device::memswap::MemSwapInfoSet;
 using mindspore::device::memswap::MemSwapManager;
 using mindspore::device::memswap::SwapKind;
 bool GPUKernelRuntime::SyncStream() { return GPUDeviceManager::GetInstance().SyncStream(stream_); }
@@ -139,6 +140,7 @@ void GPUKernelRuntime::AssignMemory(session::KernelGraph *graph) {
     InitKernelRefCount(graph);
     InitMemorySwapInfo(graph);
     InitKernelOutputAddress(graph);
+    InitKernelWorkspaceAddress(graph);
   } else {
     AssignDynamicMemory(graph);
   }
@@ -183,6 +185,56 @@ bool GPUKernelRuntime::Run(session::KernelGraph *graph) {
   return ret;
 }
 
+bool GPUKernelRuntime::SearchMemSwapScheme(const session::KernelGraph *graph) {
+  bool ret = false;
+  ClearKernelOldOutputAndWorkspace(graph);
+  if (!mem_swap_manager_->mem_swap_init()) {
+    if (!mem_swap_manager_->Init(graph)) {
+      return false;
+    }
+  }
+
+  while (!ret) {
+    if (!mem_swap_manager_->RetreatSwapInfo()) {
+      return false;
+    }
+    ret = LaunchKernelDynamic(graph, true, false);
+    if (!ret) {
+      ClearKernelOldOutputAndWorkspace(graph);
+    }
+  }
+  mem_swap_manager_->AssignHostMemory();
+
+  // Time profiling
+  ret = LaunchKernelDynamic(graph, false, true);
+  if (!ret) {
+    return ret;
+  }
+  return RefineMemSwapScheme(graph);
+}
+
+bool GPUKernelRuntime::RefineMemSwapScheme(const session::KernelGraph *graph) {
+  auto &kernels = graph->execution_order();
+  for (const auto &kernel : kernels) {
+    if (!mem_swap_manager_->QueryKernelTriggerSwapIn(kernel)) {
+      continue;
+    }
+
+    size_t swap_in_task_num = mem_swap_manager_->QueryKernelTriggerSwapInTaskNum(kernel);
+    for (size_t swap_in_task_idx = 0; swap_in_task_idx < swap_in_task_num; swap_in_task_idx++) {
+      bool ret = false;
+      while (!ret) {
+        mem_swap_manager_->AdjustSwapInPos(kernel, swap_in_task_idx);
+        ret = LaunchKernelDynamic(graph, true, false);
+        if (!ret) {
+          ClearKernelOldOutputAndWorkspace(graph);
+        }
+      }
+    }
+  }
+  return true;
+}
+
 void GPUKernelRuntime::InitKernelRefCount(const session::KernelGraph *graph) {
   MS_EXCEPTION_IF_NULL(graph);
   MemReuseUtilPtr mem_reuse_util_ptr = std::make_shared<memreuse::MemReuseUtil>();
@@ -209,6 +261,7 @@ void GPUKernelRuntime::InitMemorySwapInfo(const session::KernelGraph *graph) {
   MS_EXCEPTION_IF_NULL(mem_swap_manager);
   auto graph_id = graph->graph_id();
   mem_swap_map_[graph_id] = mem_swap_manager;
+  is_first_step_map_[graph_id] = true;
 }
 
 void GPUKernelRuntime::InitKernelOutputAddress(const session::KernelGraph *graph) {
@@ -230,6 +283,25 @@ void GPUKernelRuntime::InitKernelOutputAddress(const session::KernelGraph *graph
   }
 }
 
+void GPUKernelRuntime::InitKernelWorkspaceAddress(const session::KernelGraph *graph) {
+  MS_EXCEPTION_IF_NULL(graph);
+  auto &kernels = graph->execution_order();
+  for (const auto &kernel : kernels) {
+    auto kernel_mod = AnfAlgo::GetKernelMod(kernel);
+    MS_EXCEPTION_IF_NULL(kernel_mod);
+    auto workspace_sizes = kernel_mod->GetWorkspaceSizeList();
+    for (size_t i = 0; i < workspace_sizes.size(); ++i) {
+      auto device_address = CreateDeviceAddress(nullptr, workspace_sizes[i], "", kTypeUnknown);
+      AnfAlgo::SetWorkspaceAddr(device_address, i, kernel.get());
+    }
+  }
+}
+
+void GPUKernelRuntime::ClearKernelOldOutputAndWorkspace(const session::KernelGraph *graph) {
+  ClearKernelOutputAddress(graph);
+  ClearKernelWorkspaceAddress(graph);
+}
+
 void GPUKernelRuntime::ClearKernelOutputAddress(const session::KernelGraph *graph) {
   MS_EXCEPTION_IF_NULL(graph);
   auto &kernels = graph->execution_order();
@@ -242,6 +314,7 @@ void GPUKernelRuntime::ClearKernelOutputAddress(const session::KernelGraph *grap
         continue;
       }
       auto device_address = AnfAlgo::GetMutableOutputAddr(kernel, i, false);
+      MS_EXCEPTION_IF_NULL(device_address);
       if (device_address->ptr_) {
         mem_manager_->FreeMemFromMemPool(device_address);
       }
@@ -250,7 +323,24 @@ void GPUKernelRuntime::ClearKernelOutputAddress(const session::KernelGraph *grap
   }
 }
 
-bool GPUKernelRuntime::LaunchKernelDynamic(const session::KernelGraph *graph) {
+void GPUKernelRuntime::ClearKernelWorkspaceAddress(const session::KernelGraph *graph) {
+  MS_EXCEPTION_IF_NULL(graph);
+  auto &kernels = graph->execution_order();
+  for (const auto &kernel : kernels) {
+    auto kernel_mod = AnfAlgo::GetKernelMod(kernel);
+    MS_EXCEPTION_IF_NULL(kernel_mod);
+    auto workspace_sizes = kernel_mod->GetWorkspaceSizeList();
+    for (size_t i = 0; i < workspace_sizes.size(); ++i) {
+      auto device_address = AnfAlgo::GetMutableWorkspaceAddr(kernel, i);
+      MS_EXCEPTION_IF_NULL(device_address);
+      if (device_address->ptr_) {
+        mem_manager_->FreeMemFromMemPool(device_address);
+      }
+    }
+  }
+}
+
+bool GPUKernelRuntime::LaunchKernelDynamic(const session::KernelGraph *graph, bool mock, bool profiling) {
   MS_EXCEPTION_IF_NULL(graph);
   MS_EXCEPTION_IF_NULL(mem_reuse_util_);
   // Reset the reference count.
@@ -271,7 +361,7 @@ bool GPUKernelRuntime::LaunchKernelDynamic(const session::KernelGraph *graph) {
     if (!kernel_mod->Launch(kernel_inputs, kernel_workspaces, kernel_outputs, stream_)) {
       MS_LOG(EXCEPTION) << "Launch kernel failed.";
     }
-    FreeKernelDynamicRes(kernel, kernel_workspaces);
+    FreeKernelDynamicRes(kernel);
     UpdateMemorySwapTask(kernel);
   }
   CHECK_OP_RET_WITH_EXCEPT(SyncStream(), "SyncStream failed.");
@@ -279,13 +369,39 @@ bool GPUKernelRuntime::LaunchKernelDynamic(const session::KernelGraph *graph) {
   return true;
 }
 
+void GPUKernelRuntime::LaunchKernelWithTimeProfiling(const AnfNodePtr &kernel, const AddressPtrList &inputs,
+                                                     const AddressPtrList &workspace, const AddressPtrList &outputs) {
+  auto kernel_mod = AnfAlgo::GetKernelMod(kernel);
+  MS_EXCEPTION_IF_NULL(kernel_mod);
+  float cost_time = 0;
+  DeviceEvent start = nullptr;
+  DeviceEvent end = nullptr;
+  CHECK_OP_RET_WITH_EXCEPT(CudaDriver::CreateEvent(&start), "Failed to create event.");
+  CHECK_OP_RET_WITH_EXCEPT(CudaDriver::CreateEvent(&end), "Failed to create event.");
+
+  CHECK_OP_RET_WITH_EXCEPT(CudaDriver::RecordEvent(start, stream_), "Failed to record event to stream.");
+  CHECK_OP_RET_WITH_EXCEPT(kernel_mod->Launch(inputs, workspace, outputs, stream_), "Launch kernel failed.");
+  CHECK_OP_RET_WITH_EXCEPT(CudaDriver::RecordEvent(end, stream_), "Failed to record event to stream.");
+
+  CHECK_OP_RET_WITH_EXCEPT(CudaDriver::SyncEvent(start), "Failed to sync event.");
+  CHECK_OP_RET_WITH_EXCEPT(CudaDriver::SyncEvent(end), "Failed to sync event.");
+  CHECK_OP_RET_WITH_EXCEPT(CudaDriver::ElapsedTime(&cost_time, start, end), "Failed to record elapsed time.");
+
+  mem_swap_manager_->AddKernelExecutionPerform(kernel, cost_time);
+
+  CHECK_OP_RET_WITH_EXCEPT(CudaDriver::DestroyEvent(start), "Failed to destroy event.");
+  CHECK_OP_RET_WITH_EXCEPT(CudaDriver::DestroyEvent(end), "Failed to destroy event.");
+}
+
 bool GPUKernelRuntime::AddMemorySwapTask(const AnfNodePtr &kernel) {
   MS_EXCEPTION_IF_NULL(mem_swap_manager_);
-  auto &mem_swap_info_list = mem_swap_manager_->QueryKernelMemSwapInfo(kernel);
+  const MemSwapInfoSet &mem_swap_info_set = mem_swap_manager_->QueryKernelMemSwapInfo(kernel);
-  for (auto &mem_swap_info : mem_swap_info_list) {
+  for (auto &mem_swap_info : mem_swap_info_set) {
-    auto &kernel_exec_info = mem_swap_manager_->SearchKernelExecutionInfo(mem_swap_info.kernel_);
+    auto need_swap_kernel = mem_swap_manager_->QueryKerneByTopoOrder(mem_swap_info.topo_order_);
-    const HostAddress &host_address = kernel_exec_info.host_addrs_[mem_swap_info.output_idx_];
+    MS_EXCEPTION_IF_NULL(need_swap_kernel);
-    auto device_address = AnfAlgo::GetMutableOutputAddr(mem_swap_info.kernel_, mem_swap_info.output_idx_, false);
+    const HostAddress &host_address =
+      mem_swap_manager_->QueryKernelHostAddr(need_swap_kernel, mem_swap_info.output_idx_);
+    auto device_address = AnfAlgo::GetMutableOutputAddr(need_swap_kernel, mem_swap_info.output_idx_, false);
 
     if (mem_swap_info.swap_kind_ == SwapKind::kDeviceToHost) {
       mem_swap_manager_->AddMemSwapTask(SwapKind::kDeviceToHost, device_address, host_address);
@@ -309,9 +425,11 @@ bool GPUKernelRuntime::AddMemorySwapTask(const AnfNodePtr &kernel) {
 
 bool GPUKernelRuntime::UpdateMemorySwapInfo(const session::KernelGraph *graph) {
   MS_EXCEPTION_IF_NULL(mem_swap_manager_);
-  ClearKernelOutputAddress(graph);
+  ClearKernelOldOutputAndWorkspace(graph);
   if (!mem_swap_manager_->mem_swap_init()) {
-    mem_swap_manager_->Init(graph);
+    if (!mem_swap_manager_->Init(graph)) {
+      return false;
+    }
   }
   return mem_swap_manager_->RetreatSwapInfo();
 }
@@ -408,29 +526,6 @@ bool GPUKernelRuntime::AttemptMallocMem(const DeviceAddressPtr &device_address,
   return true;
 }
 
-void *GPUKernelRuntime::AttemptMallocMem(size_t size) {
-  MS_EXCEPTION_IF_NULL(mem_manager_);
-  MS_EXCEPTION_IF_NULL(mem_swap_manager_);
-  auto device_ptr = mem_manager_->MallocMemFromMemPool(size);
-  if (!device_ptr) {
-    if (!mem_swap_manager_->trigger_swap()) {
-      return nullptr;
-    }
-    mem_swap_manager_->SyncMemCopyStream(SwapKind::kDeviceToHost);
-    while (auto device_address_swap_out = mem_swap_manager_->UpdateSwapQueue(SwapKind::kDeviceToHost)) {
-      if (!mem_swap_manager_->FindInSwapInBlackList(device_address_swap_out->ptr_) && device_address_swap_out->ptr_) {
-        device_address_swap_out->set_status(DeviceAddressStatus::kInHost);
-        mem_manager_->FreeMemFromMemPool(device_address_swap_out);
-      }
-    }
-    device_ptr = mem_manager_->MallocMemFromMemPool(size);
-    if (!device_ptr) {
-      return nullptr;
-    }
-  }
-  return device_ptr;
-}
-
 bool GPUKernelRuntime::AllocKernelDynamicRes(const mindspore::kernel::KernelMod &kernel_mod,
                                              const mindspore::AnfNodePtr &kernel, AddressPtrList *kernel_inputs,
                                              AddressPtrList *kernel_workspaces, AddressPtrList *kernel_outputs) {
@@ -504,13 +599,13 @@ bool GPUKernelRuntime::AllocKernelWorkspaceDynamicRes(const mindspore::kernel::K
       kernel_workspaces->emplace_back(nullptr);
       continue;
     }
-    auto device_ptr = AttemptMallocMem(workspace_sizes[i]);
+    auto device_address = AnfAlgo::GetMutableWorkspaceAddr(kernel, i);
-    if (!device_ptr) {
+    if (device_address->ptr_ == nullptr && !AttemptMallocMem(device_address, workspace_sizes[i])) {
       return false;
     }
     kernel::AddressPtr workspace = std::make_shared<kernel::Address>();
     MS_EXCEPTION_IF_NULL(workspace);
-    workspace->addr = device_ptr;
+    workspace->addr = device_address->ptr_;
     workspace->size = workspace_sizes[i];
     kernel_workspaces->emplace_back(workspace);
   }
@@ -606,8 +701,7 @@ void GPUKernelRuntime::AllocCommunicationOpMemory(bool is_need_alloc_memory, boo
   }
 }
 
-void GPUKernelRuntime::FreeKernelDynamicRes(const mindspore::AnfNodePtr &kernel,
+void GPUKernelRuntime::FreeKernelDynamicRes(const mindspore::AnfNodePtr &kernel) {
-                                            const AddressPtrList &kernel_workspaces) {
   MS_EXCEPTION_IF_NULL(kernel);
   MS_EXCEPTION_IF_NULL(mem_manager_);
   MS_EXCEPTION_IF_NULL(mem_reuse_util_);
@@ -652,12 +746,13 @@ void GPUKernelRuntime::FreeKernelDynamicRes(const mindspore::AnfNodePtr &kernel,
     }
   }
   // Free the workspace of kernel.
-  for (size_t i = 0; i < kernel_workspaces.size(); ++i) {
+  auto kernel_mod = AnfAlgo::GetKernelMod(kernel);
-    auto workspace = kernel_workspaces[i];
+  MS_EXCEPTION_IF_NULL(kernel_mod);
-    if (workspace != nullptr) {
+  for (size_t i = 0; i < kernel_mod->GetWorkspaceSizeList().size(); ++i) {
-      MS_EXCEPTION_IF_NULL(workspace->addr);
+    auto device_address = AnfAlgo::GetMutableWorkspaceAddr(kernel, i);
-      mem_manager_->FreeMemFromMemPool(workspace->addr);
+    MS_EXCEPTION_IF_NULL(device_address);
-      workspace->addr = nullptr;
+    if (device_address->ptr_) {
+      mem_manager_->FreeMemFromMemPool(device_address);
     }
   }
 }

mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.h

@@ -53,11 +53,17 @@ class GPUKernelRuntime : public KernelRuntime {
   // The related functions and members for using dynamic memory pool.
   void InitKernelRefCount(const session::KernelGraph *graph);
   void InitKernelOutputAddress(const session::KernelGraph *graph);
+  void InitKernelWorkspaceAddress(const session::KernelGraph *graph);
   void InitMemorySwapInfo(const session::KernelGraph *graph);
   void ClearKernelOutputAddress(const session::KernelGraph *graph);
-  bool LaunchKernelDynamic(const session::KernelGraph *graph);
+  void ClearKernelWorkspaceAddress(const session::KernelGraph *graph);
+  void ClearKernelOldOutputAndWorkspace(const session::KernelGraph *graph);
+  bool SearchMemSwapScheme(const session::KernelGraph *graph);
+  bool RefineMemSwapScheme(const session::KernelGraph *graph);
+  bool LaunchKernelDynamic(const session::KernelGraph *graph, bool mock = false, bool profiling = false);
+  void LaunchKernelWithTimeProfiling(const AnfNodePtr &kernel, const AddressPtrList &inputs,
+                                     const AddressPtrList &workspace, const AddressPtrList &outputs);
   bool AttemptMallocMem(const DeviceAddressPtr &device_address, size_t size);
-  void *AttemptMallocMem(size_t size);
   bool AllocKernelDynamicRes(const mindspore::kernel::KernelMod &kernel_mod, const mindspore::AnfNodePtr &kernel,
                              AddressPtrList *kernel_inputs, AddressPtrList *kernel_workspaces,
                              AddressPtrList *kernel_outputs);
@@ -72,7 +78,7 @@ class GPUKernelRuntime : public KernelRuntime {
   void AllocCommunicationOpMemory(bool is_need_alloc_memory, bool is_need_free_memory,
                                   const DeviceAddressPtrList addr_list, size_t total_size,
                                   std::vector<size_t> size_list);
-  void FreeKernelDynamicRes(const mindspore::AnfNodePtr &kernel, const AddressPtrList &kernel_workspaces);
+  void FreeKernelDynamicRes(const mindspore::AnfNodePtr &kernel);
   bool AddMemorySwapTask(const AnfNodePtr &kernel);
   bool UpdateMemorySwapInfo(const session::KernelGraph *graph);
   bool UpdateMemorySwapTask(const AnfNodePtr &kernel);
@@ -81,6 +87,7 @@ class GPUKernelRuntime : public KernelRuntime {
   void ClearSwapQueue();
   std::unordered_map<uint32_t, MemReuseUtilPtr> mem_reuse_util_map_;
   std::unordered_map<uint32_t, MemSwapManagerPtr> mem_swap_map_;
+  std::unordered_map<uint32_t, bool> is_first_step_map_;
   MemReuseUtilPtr mem_reuse_util_{nullptr};
   MemSwapManagerPtr mem_swap_manager_{nullptr};
 };

mindspore/ccsrc/runtime/device/kernel_info.cc

@@ -73,6 +73,14 @@ DeviceAddress *KernelInfo::GetWorkspaceAddr(size_t index) const {
   return workspace_address_list_[index].get();
 }
 
+DeviceAddressPtr KernelInfo::GetMutableWorkspaceAddr(size_t index) const {
+  if (index >= workspace_address_list_.size()) {
+    MS_LOG(ERROR) << "Index [" << index << "] out of range";
+    return nullptr;
+  }
+  return workspace_address_list_[index];
+}
+
 bool KernelInfo::SetWorkspaceAddr(const DeviceAddressPtr &output_address, size_t index) {
   if (workspace_address_list_.empty()) {
     // parameter and valuenode

mindspore/ccsrc/runtime/device/kernel_info.h

@@ -54,6 +54,7 @@ class KernelInfo : public KernelInfoDevice {
   bool OutputAddrExist(size_t index) const;
   bool SetOutputAddr(const DeviceAddressPtr &output_address, size_t index);
   DeviceAddress *GetWorkspaceAddr(size_t index) const;
+  DeviceAddressPtr GetMutableWorkspaceAddr(size_t index) const;
   bool SetWorkspaceAddr(const DeviceAddressPtr &output_address, size_t index);
   void set_kernel_mod(const kernel::KernelModPtr &kernel_mod);
   kernel::KernelMod *MutableKernelMod() const;