Memory Leaks when train transformer model.

[gongweibao]

Background:

I found memory leaks in the process of run transformer model. Memory increases by speed about 100KB/batch. Both trainer and pserver meet the problem.

Generally, memory increases by two reasons:

• Malloced(newed) memory is not freed.
• Memory fragment

And I found two location of not freed memory use pprof tool to run all C++ unit tests:

• Fluid bytecode c++ executor. #10220
• Fix beam_search memory leak. #10185

But the memory increases over time even I solved the above.

Analysis

First, I use pprof and Valgrind to detect when run python interface, but it contains a lot of warnings

• I write a C++ executor to friendly to memory check tool, but I found nothing except the initialize memory leak.
• I compiled the debug version python for memory check tool, and the result is similar to the above.

Second, I think maybe there's memory fragment in Glibc memory pool:

• Use malloc_trim to release not used memory: it's not helpful.
• Use LD_PRELOAD tcmalloc.so, and set TCMALLOC_RELEASE_RATE=10.0(max value): it's not helpful.
• Link tcmalloc to paddle: because our complicated dependency and the dependency order, I meet free invalid pointer error and so fail to link.

Third, I think it's maybe the python memory leak:

• Use gc.collect gc.garbage to find uncollectable objects: there's nothing.

Third, I use mallinfo to trace memory consumption. It does consume some memory every batch, but I can't locate the operator: every operator use new or malloc to allocate memory in our code or std:: STL code.

Conclusion: Need help

Maybe memory fragment is the reason of memory leak. We can use malloc hook to manage our memory.

Reference:

• Understanding glibc malloc
• TCMalloc : Thread-Caching Malloc
• mallinfo
• How to use valgrind with python?
  

valgrind + debug version python:

7881 ==22452== 616,960 (63,240 direct, 553,720 indirect) bytes in 527 blocks are definitely lost in loss record 13,934 of 13,972
7882 ==22452==    at 0x4C2E0EF: operator new(unsigned long) (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
7883 ==22452==    by 0x3D549AA9: google::protobuf::internal::GenericTypeHandler<paddle::framework::proto::OpDesc>::NewFromPrototype(paddle::framework::proto::OpDesc const*,      google::protobuf::Arena*) [clone .isra.186] (in /paddle/build/release_gpu/python/paddle/fluid/core.so)
7884 ==22452==    by 0x3D5509C0: paddle::framework::proto::BlockDesc::UnsafeMergeFrom(paddle::framework::proto::BlockDesc const&) (in /paddle/build/release_gpu/python/paddle     /fluid/core.so)
7885 ==22452==    by 0x3D550D86: paddle::framework::proto::ProgramDesc::UnsafeMergeFrom(paddle::framework::proto::ProgramDesc const&) (in /paddle/build/release_gpu/python/pa     ddle/fluid/core.so)
7886 ==22452==    by 0x3C7A288D: paddle::framework::ProgramDesc::ProgramDesc(paddle::framework::proto::ProgramDesc const&) (in /paddle/build/release_gpu/python/paddle/fluid/     core.so)
7887 ==22452==    by 0x3C6F3F61: void pybind11::cpp_function::initialize<paddle::pybind::pybind11_init()::{lambda(paddle::framework::ProgramDesc const&, std::vector<std::arr     ay<unsigned long, 2ul>, std::allocator<std::array<unsigned long, 2ul> > > const&)#39}, paddle::framework::ProgramDesc*, paddle::framework::ProgramDesc const&, std::vect     or<std::array<unsigned long, 2ul>, std::allocator<std::array<unsigned long, 2ul> > > const&, pybind11::name, pybind11::scope, pybind11::sibling>(paddle::pybind::pybind1     1_init()::{lambda(paddle::framework::ProgramDesc const&, std::vector<std::array<unsigned long, 2ul>, std::allocator<std::array<unsigned long, 2ul> > > const&)#39}&&, pa     ddle::framework::ProgramDesc* (*)(paddle::framework::ProgramDesc const&, std::vector<std::array<unsigned long, 2ul>, std::allocator<std::array<unsigned long, 2ul> > > c     onst&), pybind11::name const&, pybind11::scope const&, pybind11::sibling const&)::{lambda(pybind11::detail::function_call&)#3}::_FUN(pybind11::detail::function_call) (i     n /paddle/build/release_gpu/python/paddle/fluid/core.so)
7888 ==22452==    by 0x3C713823: pybind11::cpp_function::dispatcher(_object*, _object*, _object*) (in /paddle/build/release_gpu/python/paddle/fluid/core.so)
7889 ==22452==    by 0x4BC3F9: PyEval_EvalFrameEx (in /usr/bin/python2.7)
7890 ==22452==    by 0x4B9AB5: PyEval_EvalCodeEx (in /usr/bin/python2.7)
7891 ==22452==    by 0x4C1E6E: PyEval_EvalFrameEx (in /usr/bin/python2.7)
7892 ==22452==    by 0x4B9AB5: PyEval_EvalCodeEx (in /usr/bin/python2.7)
7893 ==22452==    by 0x4C16E6: PyEval_EvalFrameEx (in /usr/bin/python2.7)
.....
7938 ==22452==    definitely lost: 264,370 bytes in 1,939 blocks
7939 ==22452==    indirectly lost: 799,257 bytes in 16,449 blocks
7940 ==22452==      possibly lost: 6,991,543 bytes in 49,725 blocks
7941 ==22452==    still reachable: 247,616,073 bytes in 535,440 blocks
7942 ==22452==                       of which reachable via heuristic:
7943 ==22452==                         stdstring          : 7,891 bytes in 122 blocks
7944 ==22452==         suppressed: 0 bytes in 0 blocks
7945 ==22452== Reachable blocks (those to which a pointer was found) are not shown.
7946 ==22452== To see them, rerun with: --leak-check=full --show-leak-kinds=all

glibc malloc trace：

Diff is the memory consumed by operator.

Diff is the memory consumed by executor of every batch.

[gongweibao]

And I can't figure out why this #10358 resolved the inference memory leak.