set the max size of cudapinned memory

paddle/fluid/memory/detail/system_allocator.cc

@@ -125,11 +125,11 @@ bool GPUAllocator::UseGpu() const { return true; }
 void* CUDAPinnedAllocator::Alloc(size_t& index, size_t size) {
   if (size <= 0) return nullptr;
 
-  // NOTE: here, we use CpuMaxAllocSize()/2 as the maximum memory size
+  // NOTE: here, we use CUDAPinnedMaxAllocSize as the maximum memory size
   // of host pinned allocation. Allocates too much would reduce
   // the amount of memory available to the underlying system for paging.
   size_t usable =
-      paddle::platform::CpuMaxAllocSize() / 2 - cuda_pinnd_alloc_size_;
+      paddle::platform::CUDAPinnedMaxAllocSize() - cuda_pinnd_alloc_size_;
 
   if (size > usable) return nullptr;
 

paddle/fluid/memory/memory.cc

@@ -116,8 +116,8 @@ BuddyAllocator* GetCUDAPinnedBuddyAllocator() {
   static BuddyAllocator* ba = NULL;
   if (ba == NULL) {
     ba = new BuddyAllocator(new detail::CUDAPinnedAllocator,
-                            platform::CpuMinChunkSize(),
-                            platform::CpuMaxChunkSize());
+                            platform::CUDAPinnedMinChunkSize(),
+                            platform::CUDAPinnedMaxChunkSize());
   }
   return ba;
 }

paddle/fluid/memory/memory_test.cc

@@ -143,7 +143,7 @@ TEST(BuddyAllocator, GPUMultAlloc) {
 
 size_t align(size_t size, paddle::platform::CUDAPinnedPlace place) {
   size += sizeof(paddle::memory::detail::Metadata);
-  size_t alignment = paddle::platform::CpuMinChunkSize();
+  size_t alignment = paddle::platform::CUDAPinnedMinChunkSize();
   size_t remaining = size % alignment;
   return remaining == 0 ? size : size + (alignment - remaining);
 }

paddle/fluid/platform/cpu_info.cc

@@ -27,6 +27,10 @@ DEFINE_double(fraction_of_cpu_memory_to_use, 1,
               "Default use 100% of CPU memory for PaddlePaddle,"
               "reserve the rest for page tables, etc");
 
+DEFINE_double(fraction_of_cuda_pinned_memory_to_use, 0.5,
+              "Default use 100% of CPU memory for PaddlePaddle,"
+              "reserve the rest for page tables, etc");
+
 namespace paddle {
 namespace platform {
 
@@ -62,5 +66,22 @@ size_t CpuMaxChunkSize() {
   return CpuMaxAllocSize() / 32;
 }
 
+size_t CUDAPinnedMaxAllocSize() {
+  // For distributed systems, it requires configuring and limiting
+  // the fraction of memory to use.
+  return FLAGS_fraction_of_cuda_pinned_memory_to_use * CpuTotalPhysicalMemory();
+}
+
+size_t CUDAPinnedMinChunkSize() {
+  // Allow to allocate the minimum chunk size is 64 KB.
+  return 1 << 16;
+}
+
+size_t CUDAPinnedMaxChunkSize() {
+  // Allow to allocate the maximum chunk size is roughly 0.39% of CUDA_PINNED
+  // memory.
+  return CUDAPinnedMaxAllocSize() / 256;
+}
+
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/cpu_info.h

@@ -22,11 +22,20 @@ namespace platform {
 //! Get the maximum allocation size for a machine.
 size_t CpuMaxAllocSize();
 
+//! Get the maximum allocation size for a machine.
+size_t CUDAPinnedMaxAllocSize();
+
 //! Get the minimum chunk size for buddy allocator.
 size_t CpuMinChunkSize();
 
 //! Get the maximum chunk size for buddy allocator.
 size_t CpuMaxChunkSize();
 
+//! Get the minimum chunk size for buddy allocator.
+size_t CUDAPinnedMinChunkSize();
+
+//! Get the maximum chunk size for buddy allocator.
+size_t CUDAPinnedMaxChunkSize();
+
 }  // namespace platform
 }  // namespace paddle