Example of restricting part of graph to run on single core

gistfile1.txt

# try running cpu intensive test on two devices

import tensorflow as tf
import time

def matmul_op():
  """Multiply two matrices together"""
  
  n = 2000
  a = tf.ones((n, n), dtype=tf.float32)
  return tf.matmul(a, a)/n

slow_op  = matmul_op

with tf.device("/cpu:0"):
    one = slow_op()
with tf.device("/cpu:1"):
    another_one = slow_op()

config = tf.ConfigProto(device_count={"CPU": 2},
                        inter_op_parallelism_threads=2,
                        intra_op_parallelism_threads=1)
config.graph_options.optimizer_options.opt_level = -1

sess = tf.Session(config=config)
two = one+another_one

# pre-warm the kernels
sess.run(one)

start  = time.time()
sess.run(one)
elapsed_time = time.time() - start
print("Single op: %2.4f sec"%(elapsed_time))

start  = time.time()
sess.run(two)
elapsed_time2 = time.time()-start
print("Two ops in parallel: %.2f sec (%.2f times slower)"%(elapsed_time2,
      elapsed_time2/elapsed_time))

Hi yaroslavvb,
why do we need to pre-warm the kernel? I think it is for more accurate result. If it is so, what is the reason behind it?

I multi session mean multi threading ?
is each one session run in different thread ?