!8291 [MD][GPU] minddata use pin memory can improve efficiency

From: @xiefangqi
Reviewed-by: @pandoublefeng,@jonyguo
Signed-off-by: @jonyguo

mindspore/ccsrc/backend/kernel_compiler/gpu/data/dataset_init_kernel.h

@@ -49,7 +49,7 @@ class DatasetInitKernel : public GpuKernel {
   std::vector<size_t> workspace_size_list_;
 
   // The capacity of buffer Q.
-  size_t buffer_q_capacity_{1};
+  size_t buffer_q_capacity_{2};
 };
 
 MS_REG_GPU_KERNEL(InitDataSetQueue, DatasetInitKernel)

mindspore/ccsrc/minddata/dataset/engine/cache/CMakeLists.txt

@@ -42,6 +42,7 @@ if (ENABLE_CACHE)
       storage_container.cc)
 
   add_executable(cache_server cache_main.cc)
+  if (ENABLE_GPU)
     target_link_libraries(cache_server
         engine-cache-server
         _c_dataengine
@@ -49,9 +50,22 @@ if (ENABLE_CACHE)
         mindspore::protobuf
         mindspore::grpc++
         mindspore_gvar
+        ${CUDNN_LIBRARY_PATH}
         ${PYTHON_LIBRARIES}
         ${SECUREC_LIBRARY}
         pthread)
+  else()
+    target_link_libraries(cache_server
+        engine-cache-server
+        _c_dataengine
+        _c_mindrecord
+        mindspore::protobuf
+        mindspore::grpc++
+        mindspore_gvar
+        ${PYTHON_LIBRARIES}
+        ${SECUREC_LIBRARY}
+        pthread)
+  endif()
 
   if (USE_GLOG)
     target_link_libraries(cache_server mindspore::glog)

mindspore/ccsrc/minddata/dataset/engine/datasetops/device_queue_op.cc

@@ -91,7 +91,7 @@ Status DeviceQueueOp::operator()() {
 #endif
   } else if (device_type_ == DeviceType::GPU) {
 #ifdef ENABLE_GPUQUE
-    RETURN_IF_NOT_OK(CircularPool::CreateCircularPool(&pool_));
+    RETURN_IF_NOT_OK(CircularPool::CreateCircularPool(&pool_, -1, 1024, false, true));
     RETURN_IF_NOT_OK(SendDataToGPU());
 #endif
   } else if (device_type_ == DeviceType::CPU) {

mindspore/ccsrc/minddata/dataset/util/arena.cc

@@ -235,14 +235,43 @@ std::ostream &operator<<(std::ostream &os, const ArenaImpl &s) {
 Status Arena::Init() {
   try {
     int64_t sz = size_in_MB_ * 1048576L;
+#ifdef ENABLE_GPUQUE
+    if (is_cuda_malloc_) {
+      auto ret = cudaHostAlloc(&ptr_, sz, cudaHostAllocDefault);
+      if (ret != cudaSuccess) {
+        MS_LOG(ERROR) << "cudaHostAlloc failed, ret[" << static_cast<int>(ret) << "], " << cudaGetErrorString(ret);
+        return Status(StatusCode::kOutOfMemory);
+      }
+      impl_ = std::make_unique<ArenaImpl>(ptr_, sz);
+    } else {
       RETURN_IF_NOT_OK(DeMalloc(sz, &ptr_, false));
       impl_ = std::make_unique<ArenaImpl>(ptr_, sz);
+    }
+#else
+    RETURN_IF_NOT_OK(DeMalloc(sz, &ptr_, false));
+    impl_ = std::make_unique<ArenaImpl>(ptr_, sz);
+#endif
   } catch (std::bad_alloc &e) {
     return Status(StatusCode::kOutOfMemory);
   }
   return Status::OK();
 }
 
+#ifdef ENABLE_GPUQUE
+Arena::Arena(size_t val_in_MB, bool is_cuda_malloc)
+    : ptr_(nullptr), size_in_MB_(val_in_MB), is_cuda_malloc_(is_cuda_malloc) {}
+
+Status Arena::CreateArena(std::shared_ptr<Arena> *p_ba, size_t val_in_MB, bool is_cuda_malloc) {
+  RETURN_UNEXPECTED_IF_NULL(p_ba);
+  auto ba = new (std::nothrow) Arena(val_in_MB, is_cuda_malloc);
+  if (ba == nullptr) {
+    return Status(StatusCode::kOutOfMemory);
+  }
+  (*p_ba).reset(ba);
+  RETURN_IF_NOT_OK(ba->Init());
+  return Status::OK();
+}
+#else
 Arena::Arena(size_t val_in_MB) : ptr_(nullptr), size_in_MB_(val_in_MB) {}
 
 Status Arena::CreateArena(std::shared_ptr<Arena> *p_ba, size_t val_in_MB) {
@@ -255,5 +284,6 @@ Status Arena::CreateArena(std::shared_ptr<Arena> *p_ba, size_t val_in_MB) {
   RETURN_IF_NOT_OK(ba->Init());
   return Status::OK();
 }
+#endif
 }  // namespace dataset
 }  // namespace mindspore

mindspore/ccsrc/minddata/dataset/util/arena.h

@@ -22,6 +22,9 @@
 #include "minddata/dataset/util/allocator.h"
 #include "minddata/dataset/util/memory_pool.h"
 #include "minddata/dataset/util/treap.h"
+#ifdef ENABLE_GPUQUE
+#include <cuda_runtime_api.h>
+#endif
 
 #define ARENA_LOG_BLK_SZ (6u)
 #define ARENA_BLK_SZ (static_cast<uint16_t>(1u << ARENA_LOG_BLK_SZ))
@@ -105,10 +108,18 @@ class Arena : public MemoryPool {
   Arena(const Arena &) = delete;
   Arena &operator=(const Arena &) = delete;
   ~Arena() override {
+#ifdef ENABLE_GPUQUE
+    if (is_cuda_malloc_) {
+      if (ptr_) {
+        (void)cudaFreeHost(ptr_);
+      }
+    }
+#else
     if (ptr_ != nullptr) {
       free(ptr_);
     }
     ptr_ = nullptr;
+#endif
   }
 
   /// As a derived class of MemoryPool, we have to implement the following.
@@ -140,16 +151,27 @@ class Arena : public MemoryPool {
     return os;
   }
 
+#ifdef ENABLE_GPUQUE
+  /// The only method to create an arena.
+  static Status CreateArena(std::shared_ptr<Arena> *p_ba, size_t val_in_MB = 4096, bool is_cuda_malloc = false);
+#else
   /// The only method to create an arena.
   static Status CreateArena(std::shared_ptr<Arena> *p_ba, size_t val_in_MB = 4096);
+#endif
 
  protected:
   mutable std::mutex mux_;
   std::unique_ptr<ArenaImpl> impl_;
   void *ptr_;
   size_t size_in_MB_;
+#ifdef ENABLE_GPUQUE
+  bool is_cuda_malloc_;
+
+  explicit Arena(size_t val_in_MB = 4096, bool is_cuda_malloc = false);
+#else
 
   explicit Arena(size_t val_in_MB = 4096);
+#endif
 
   Status Init();
 };

mindspore/ccsrc/minddata/dataset/util/circular_pool.cc

@@ -27,7 +27,11 @@ namespace dataset {
 Status CircularPool::AddOneArena() {
   Status rc;
   std::shared_ptr<Arena> b;
+#ifdef ENABLE_GPUQUE
+  RETURN_IF_NOT_OK(Arena::CreateArena(&b, arena_size_, is_cuda_malloc_));
+#else
   RETURN_IF_NOT_OK(Arena::CreateArena(&b, arena_size_));
+#endif
   tail_ = b.get();
   cur_size_in_mb_ += arena_size_;
   mem_segments_.push_back(std::move(b));
@@ -194,12 +198,43 @@ int CircularPool::PercentFree() const {
   }
 }
 
+#ifdef ENABLE_GPUQUE
+CircularPool::CircularPool(int max_size_in_gb, int arena_size, bool is_cuda_malloc)
+    : unlimited_(max_size_in_gb <= 0),
+      max_size_in_mb_(unlimited_ ? std::numeric_limits<int32_t>::max() : max_size_in_gb * 1024),
+      arena_size_(arena_size),
+      is_cuda_malloc_(is_cuda_malloc),
+      cur_size_in_mb_(0) {}
+#else
 CircularPool::CircularPool(int max_size_in_gb, int arena_size)
     : unlimited_(max_size_in_gb <= 0),
       max_size_in_mb_(unlimited_ ? std::numeric_limits<int32_t>::max() : max_size_in_gb * 1024),
       arena_size_(arena_size),
       cur_size_in_mb_(0) {}
+#endif
 
+#ifdef ENABLE_GPUQUE
+Status CircularPool::CreateCircularPool(std::shared_ptr<MemoryPool> *out_pool, int max_size_in_gb, int arena_size,
+                                        bool createOneArena, bool is_cuda_malloc) {
+  Status rc;
+  if (out_pool == nullptr) {
+    RETURN_STATUS_UNEXPECTED("pPool is null");
+  }
+  auto pool = new (std::nothrow) CircularPool(max_size_in_gb, arena_size, is_cuda_malloc);
+  if (pool == nullptr) {
+    return Status(StatusCode::kOutOfMemory);
+  }
+  if (createOneArena) {
+    rc = pool->AddOneArena();
+  }
+  if (rc.IsOk()) {
+    (*out_pool).reset(pool);
+  } else {
+    delete pool;
+  }
+  return rc;
+}
+#else
 Status CircularPool::CreateCircularPool(std::shared_ptr<MemoryPool> *out_pool, int max_size_in_gb, int arena_size,
                                         bool createOneArena) {
   Status rc;
@@ -220,6 +255,7 @@ Status CircularPool::CreateCircularPool(std::shared_ptr<MemoryPool> *out_pool, i
   }
   return rc;
 }
+#endif
 
 CircularPool::~CircularPool() = default;
 }  // namespace dataset

mindspore/ccsrc/minddata/dataset/util/circular_pool.h

@@ -85,8 +85,13 @@ class CircularPool : public MemoryPool {
     return os;
   }
 
+#ifdef ENABLE_GPUQUE
+  static Status CreateCircularPool(std::shared_ptr<MemoryPool> *out_pool, int max_size_in_gb = -1,
+                                   int arena_size = 4096, bool create_one_arena = false, bool is_cuda_malloc = false);
+#else
   static Status CreateCircularPool(std::shared_ptr<MemoryPool> *out_pool, int max_size_in_gb = -1,
                                    int arena_size = 4096, bool create_one_arena = false);
+#endif
 
  private:
   ListOfArenas mem_segments_;
@@ -96,9 +101,16 @@ class CircularPool : public MemoryPool {
   int arena_size_;
   int cur_size_in_mb_;
   RWLock rw_lock_;
+#ifdef ENABLE_GPU
+  bool is_cuda_malloc_;
+
+  // We can take negative or 0 as input which means unlimited.
+  CircularPool(int max_size_in_gb, int arena_size, bool is_cuda_malloc);
+#else
 
   // We can take negative or 0 as input which means unlimited.
   CircularPool(int max_size_in_gb, int arena_size);
+#endif
 
   Status AddOneArena();
 };