tensorflow: Bug: tf.Variable uses always twice the memory (on the CPU)

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 16.04
• TensorFlow installed from (source or binary): tested on both
• TensorFlow version (use command below): 1.3 for pip / 0cfb16e025b3d20e8c8aca431fc0887814817c44 for self-compiled
• Python version: Python 3.4.3 [GCC 4.9.2] on linux
• Bazel version (if compiling from source): 0.5.4
• CUDA/cuDNN version: not used
• GPU model and memory: not used

Describe the problem

Every tf.variable occupies always twice the necessary memory: Once for the tf.constant vector that is created for the initializer and once as the persistend storage. Code to reproduce:

import os
os.environ['TF_CPP_MIN_VLOG_LEVEL'] = '100' #print all
import tensorflow as tf

runs = 1
N = int(1024 * 1024 * 1.1)
M = int(1024 / 8)
print("testing allocation of {:.2f} MB".format(N*M*8. / 1024 / 1024))

#does not matter which version you use:
v = tf.Variable(tf.ones([M, N], tf.float64), name="var1")
#v = tf.get_variable(shape=(M,N), initializer=tf.ones_initializer, dtype=tf.float64, name='var1', trainable=False)

init = tf.global_variables_initializer()
for i in range(runs):
      print("start session")
      with tf.Session() as sess:
            print("start init")
            sess.run(init)

In my self-compiled version the output contains the following lines:

 tensorflow/core/common_runtime/bfc_allocator.cc:133] Extending allocation by 2.00GiB bytes.
 tensorflow/core/common_runtime/bfc_allocator.cc:137] Total allocated bytes: 4.00GiB

For the pip version, the current allocation is not displayed, but observing htop during the execution or using mprof reveals the same.

The issue seems to occur because assign does not reuse the memory of tf.ones but instead allocates additional memory. Related lines: assign_op.h context->forward_input returns null in this case, because the memory of tf.ones has a ref count of 2. (I dont know why) see op_kernel.cc input->RefCountIsOne() is therefore false.

I tried to comment out the input->RefCountIsOne() check, but then it still doesn’t work because of the output_attr.IsEqualOrLessRestrictiveThan() check. If you remove this check too, the memory usage finaly drops to the expected value, the memory of tf.ones is reused.
But this is not a real solution, because I don’t know how this would effect other operations and it seems to break the memory freeing.

I think this bug is quiet serious, because it affects nearly all computations. Is this an already known bug?