run dist demo word2vec failed

[Yancey1989]

Start up 10 pservers + 10 trainers + turn on sparse update, pserver would crash:

server-1:

I0301 08:39:03.283473 12305 executor.cc:134] CPUPlace Op(sgd), inputs:{Grad[fc_1.w_0@GRAD.block8[26, 2073]({})], LearningRate[tmp_9[1]({})], Param[fc_1.w_0.block8[26, 2073]({})]}, outputs:{ParamOut[fc_1.w_0.block8[26, 2073]({})]}.
I0301 08:39:03.283483 12305 operator.cc:521] expected_kernel_key:data_type[float32]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
I0301 08:39:03.310343 12305 listen_and_serv_op.cc:117] received grad: shared_w@GRAD.block1.trainer_9
*** Aborted at 1519893543 (unix time) try "date -d @1519893543" if you are using GNU date ***
PC: @                0x0 (unknown)
*** SIGSEGV (@0x0) received by PID 12305 (TID 0x7f4879419700) from PID 0; stack trace: ***
    @     0x7f4878df3390 (unknown)
    @     0x7f482da161e7 paddle::memory::detail::MetadataCache::load()
    @     0x7f482da15c73 paddle::memory::detail::MemoryBlock::type()
    @     0x7f482da150a1 paddle::memory::detail::BuddyAllocator::Free()
    @     0x7f482da11060 paddle::memory::Free<>()
    @     0x7f482d97321e paddle::framework::Tensor::PlaceholderImpl<>::~PlaceholderImpl()
    @     0x7f482d979a26 std::_Sp_counted_base<>::_M_release()
    @     0x7f482d97d27b paddle::framework::Tensor::mutable_data()
    @     0x7f482e0e55ae paddle::framework::TensorFromStream()
    @     0x7f482e0e0481 paddle::framework::DeserializeFromStream()
    @     0x7f482e0c3b8b paddle::operators::detail::DeserializeFromMessage()
    @     0x7f482e017ad1 paddle::operators::ListenAndServOp::RunImpl()
    @     0x7f482e0772a8 paddle::framework::OperatorBase::Run()

server-2:

E0301 08:39:03.331059 12306 listen_and_serv_op.cc:139] run sub program error enforce numel() > 0 failed, 0 <= 0
When calling this method, the Tensor's numel must be larger than zero. Please check Tensor::Resize has been called first. at [/paddle/paddle/fluid/framework/tensor_impl.h:123]
PaddlePaddle Call Stacks:
0       0x7f98da6a876cp paddle::platform::EnforceNotMet::EnforceNotMet(std::__exception_ptr::exception_ptr, char const*, int) + 572
1       0x7f98da6ae521p paddle::framework::Tensor::mutable_data(boost::variant<paddle::platform::CUDAPlace, paddle::platform::CPUPlace, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_>, std::type_index) + 1233
2       0x7f98dad0fafbp paddle::operators::SumKernel<paddle::platform::CPUDeviceContext, float>::Compute(paddle::framework::ExecutionContext const&) const + 1323
3       0x7f98dadaa944p paddle::framework::OperatorWithKernel::RunImpl(paddle::framework::Scope const&, boost::variant<paddle::platform::CUDAPlace, paddle::platform::CPUPlace, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_> const&) const + 1556
4       0x7f98dada82a8p paddle::framework::OperatorBase::Run(paddle::framework::Scope const&, boost::variant<paddle::platform::CUDAPlace, paddle::platform::CPUPlace, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_, boost::detail::variant::void_> const&) + 72

[Yancey1989]

Update progress:
There are two bugs in this issue:

1. As the logs server-2, if one of the pserver instances didn't receive any sparse vars, the GRAD var would not be initialized, we should deal with this situation.

2. As the log server-1, looks like a memory bug, but I'm not sure, will follow it.