add free memory

core/predictor/framework/infer.h

@@ -501,6 +501,7 @@ class FluidInferEngine : public CloneDBReloadableInferEngine<EngineCore> {
       paddle::PaddleBuf paddleBuf(databuf_char, databuf_size);
       tensor_out.data = paddleBuf;
       tensorVector_out_pointer->push_back(tensor_out);
+      MempoolWrapper::instance().free(databuf_char, databuf_size);
     }
     return 0;
   }

core/predictor/framework/memory.cpp

@@ -112,6 +112,23 @@ void* MempoolWrapper::malloc(size_t size) {
   return mempool->malloc(size);
 }
 
+void MempoolWrapper::free(void* p, size_t size) {
+  MempoolRegion* mempool_region =
+      (MempoolRegion*)THREAD_GETSPECIFIC(_bspec_key);
+  if (mempool_region == NULL) {
+    LOG(WARNING) << "THREAD_GETSPECIFIC() returned NULL";
+    return;
+  }
+
+  im::Mempool* mempool = mempool_region->mempool();
+  if (!mempool) {
+    LOG(WARNING) << "Cannot free memory:" << size
+                 << ", since mempool is not thread initialized";
+    return;
+  }
+  return mempool->free(p,size);
+}
+
 }  // namespace predictor
 }  // namespace paddle_serving
 }  // namespace baidu

core/predictor/framework/memory.h

@@ -38,6 +38,8 @@ class MempoolWrapper {
 
   void* malloc(size_t size);
 
+  void free(void* p, size_t size);
+
  private:
   // im::fugue::memory::Region _region;
   THREAD_KEY_T _bspec_key;

core/predictor/mempool/mempool.cpp

@@ -24,7 +24,7 @@ namespace fugue {
 namespace memory {
 
 void Region::init() {
-  _big_mem_capacity = 32 * 1024 * 1024;
+  _big_mem_capacity = 64 * 1024 * 1024;//64MB
   _big_mem_start = new char[_big_mem_capacity];
 }
 

core/predictor/mempool/mempool.h

@@ -264,7 +264,7 @@ struct BlockReference {
 // total number is 32*1024
 class BlockFreeList {
  public:
-  static const int MAX_BLOCK_COUNT = 32 * 1024;
+  static const int MAX_BLOCK_COUNT = 256 * 1024;
   typedef lockfree::FreeList<Block, MAX_BLOCK_COUNT> BlockFreeListType;
   static BlockFreeListType* instance() {
     static BlockFreeListType singleton;
@@ -340,7 +340,9 @@ class Region {
   static const int BIG_MEM_THRESHOLD =
       2 * 1024 *
       1024;  // 2MB,means when you need less than 2M, get memory from Block.
-  static const int BIGNODE_MEM_THRESHOLD = 4 * 1024 * 1024;  // 4MB
+
+  // 64MB,means when you need less than 64MB, get memory from BigMemory instead of BigNode
+  static const int BIGNODE_MEM_THRESHOLD = (64 * 1024 * 1024 + 1);  
   static const int COUNTER_SIZE =
       BIGNODE_MEM_THRESHOLD / BIG_MEM_THRESHOLD + 1;  // this is not used
 
@@ -421,7 +423,8 @@ class Mempool {
     }
 
     // 可能返回的是单独Region中malloc的内存。
-    // 也可能是Block，例如new_size=1M, old_data原本的指针头就在1.2M处，old_size =
+    // 也可能是Block，例如new_size=1M, old_data原本的指针头就在1.2M处，old_size
+    // =
     // 0.5M
     // 此时,_free_size = 0.3M，new_size<2M,但是required = 1-0.5 >0.3
     // 分配出来的就是Block，但是该Block没有并很完美的利用完全。