Merge pull request #12364 from JiayiFeng/dev_add_FLAG_free_idle_memory

add flag to prevent unnessary memory free

paddle/fluid/memory/detail/buddy_allocator.cc

@@ -15,6 +15,10 @@ limitations under the License. */
 #include "paddle/fluid/memory/detail/buddy_allocator.h"
 #include "glog/logging.h"
 
+DEFINE_bool(free_idle_memory, false,
+            "If it is true, Paddle will try to free idle memory trunks during "
+            "running time.");
+
 namespace paddle {
 namespace memory {
 namespace detail {
@@ -152,13 +156,14 @@ void BuddyAllocator::Free(void* p) {
   pool_.insert(
       IndexSizeAddress(block->index(cache_), block->total_size(cache_), block));
 
-  // Clean up if existing too much free memory
-
-  // Prefer freeing fallback allocation first
-  CleanIdleFallBackAlloc();
+  if (FLAGS_free_idle_memory) {
+    // Clean up if existing too much free memory
+    // Prefer freeing fallback allocation first
+    CleanIdleFallBackAlloc();
 
-  // Free normal allocation
-  CleanIdleNormalAlloc();
+    // Free normal allocation
+    CleanIdleNormalAlloc();
+  }
 }
 
 size_t BuddyAllocator::Used() { return total_used_; }

python/paddle/fluid/__init__.py

@@ -62,33 +62,33 @@ from paddle.fluid.layers.math_op_patch import monkey_patch_variable
 Tensor = LoDTensor
 
 __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + \
-          trainer.__all__ + inferencer.__all__ + transpiler.__all__ + \
-          parallel_executor.__all__ + lod_tensor.__all__ + [
-              'io',
-              'initializer',
-              'layers',
-              'contrib',
-              'transpiler',
-              'nets',
-              'optimizer',
-              'learning_rate_decay',
-              'backward',
-              'regularizer',
-              'LoDTensor',
-              'LoDTensorArray',
-              'CPUPlace',
-              'CUDAPlace',
-              'CUDAPinnedPlace',
-              'Tensor',
-              'ParamAttr',
-              'WeightNormParamAttr',
-              'DataFeeder',
-              'clip',
-              'profiler',
-              'unique_name',
-              'recordio_writer',
-              'Scope',
-          ]
+    trainer.__all__ + inferencer.__all__ + transpiler.__all__ + \
+    parallel_executor.__all__ + lod_tensor.__all__ + [
+        'io',
+        'initializer',
+        'layers',
+        'contrib',
+        'transpiler',
+        'nets',
+        'optimizer',
+        'learning_rate_decay',
+        'backward',
+        'regularizer',
+        'LoDTensor',
+        'LoDTensorArray',
+        'CPUPlace',
+        'CUDAPlace',
+        'CUDAPinnedPlace',
+        'Tensor',
+        'ParamAttr',
+        'WeightNormParamAttr',
+        'DataFeeder',
+        'clip',
+        'profiler',
+        'unique_name',
+        'recordio_writer',
+        'Scope',
+    ]
 
 
 def __bootstrap__():
@@ -123,7 +123,7 @@ def __bootstrap__():
     read_env_flags = [
         'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir',
         'eager_delete_scope', 'use_mkldnn', 'initial_cpu_memory_in_mb',
-        'init_allocated_mem'
+        'init_allocated_mem', 'free_idle_memory'
     ]
     if core.is_compiled_with_dist():
         read_env_flags.append('rpc_deadline')