Training a simple model in Tensorflow GPU slower than CPU

Question

I have set up a simple linear regression problem in Tensorflow, and have created simple conda environments using Tensorflow CPU and GPU both in 1.13.1 (using CUDA 10.0 in the backend on an NVIDIA Quadro P600).

However, it looks like the GPU environment always takes longer time than the CPU environment. The code I'm running is below.

import time
import warnings
import numpy as np
import scipy

import tensorflow as tf
import tensorflow_probability as tfp

from tensorflow_probability import edward2 as ed
from tensorflow.python.ops import control_flow_ops
from tensorflow_probability import distributions as tfd



# Handy snippet to reset the global graph and global session.
def reset_g():
    with warnings.catch_warnings():
        warnings.simplefilter('ignore')
        tf.reset_default_graph()
        try:
            sess.close()
        except:
            pass


N = 35000
inttest = np.ones(N).reshape(N, 1)
stddev_raw = 0.09

true_int = 1.
true_b1 = 0.15
true_b2 = 0.7

np.random.seed(69)

X1 = (np.atleast_2d(np.linspace(
    0., 2., num=N)).T).astype(np.float64)
X2 = (np.atleast_2d(np.linspace(
    2., 1., num=N)).T).astype(np.float64)
Ytest = true_int + (true_b1*X1) + (true_b2*X2) + \
    np.random.normal(size=N, scale=stddev_raw).reshape(N, 1)

Ytest = Ytest.reshape(N, )
X1 = X1.reshape(N, )
X2 = X2.reshape(N, )

reset_g()

# Create data and param
model_X1 = tf.placeholder(dtype=tf.float64, shape=[N, ])
model_X2 = tf.placeholder(dtype=tf.float64, shape=[N, ])
model_Y = tf.placeholder(dtype=tf.float64, shape=[N, ])

alpha = tf.get_variable(shape=[1], name='alpha', dtype=tf.float64)
# these two params need shape of one if using trainable distro
beta1 = tf.get_variable(shape=[1], name='beta1', dtype=tf.float64)
beta2 = tf.get_variable(shape=[1], name='beta2', dtype=tf.float64)

# Yhat
tf_pred = (tf.multiply(model_X1, beta1) + tf.multiply(model_X2, beta2) + alpha)


# # Make difference of squares
# resid = tf.square(model_Y - tf_pred)
# loss = tf.reduce_sum(resid)

# # Make a Likelihood function based on simple stuff
stddev = tf.square(tf.get_variable(shape=[1],
                                    name='stddev', dtype=tf.float64))
covar = tfd.Normal(loc=model_Y, scale=stddev)
loss = -1.0*tf.reduce_sum(covar.log_prob(tf_pred))



# Trainer
lr=0.005
N_ITER = 20000

opt = tf.train.AdamOptimizer(lr, beta1=0.95, beta2=0.95)
train = opt.minimize(loss)


with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    start = time.time()
    for step in range(N_ITER):
        out_l, out_b1, out_b2, out_a, laws = sess.run([train, beta1, beta2, alpha, loss],
                                                  feed_dict={model_X1: X1,
                                                             model_X2: X2,
                                                             model_Y: Ytest})

        if step % 500 == 0:
            print('Step: {s}, loss = {l}, alpha = {a:.3f}, beta1 = {b1:.3f}, beta2 = {b2:.3f}'.format(
                s=step, l=laws, a=out_a[0], b1=out_b1[0], b2=out_b2[0]))
    print(f"True: alpha = {true_int}, beta1 = {true_b1}, beta2 = {true_b2}")
    end = time.time()
    print(end-start)

Here are some outputs printed if they're any indicative of what's happening:

For the CPU run:

Colocations handled automatically by placer.
2019-04-18 09:00:56.329669: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2019-04-18 09:00:56.351151: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2904000000 Hz
2019-04-18 09:00:56.351672: I tensorflow/compiler/xla/service/service.cc:150] XLA service 0x558fefe604c0 executing computations on platform Host. Devices:
2019-04-18 09:00:56.351698: I tensorflow/compiler/xla/service/service.cc:158]   StreamExecutor device (0): <undefined>, <undefined>

For the GPU run:

Instructions for updating:
Call initializer instance with the dtype argument instead of passing it to the constructor
W0418 09:03:21.674947 139956864096064 deprecation.py:506] From /home/sadatnfs/.conda/envs/tf_gpu/lib/python3.6/site-packages/tensorflow/python/training/slot_creator.py:187: calling Zeros.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.
Instructions for updating:
Call initializer instance with the dtype argument instead of passing it to the constructor
2019-04-18 09:03:21.712913: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2019-04-18 09:03:21.717598: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcuda.so.1
2019-04-18 09:03:21.951277: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:1009] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2019-04-18 09:03:21.952212: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x55e583bc4480 executing computations on platform CUDA. Devices:
2019-04-18 09:03:21.952225: I tensorflow/compiler/xla/service/service.cc:175]   StreamExecutor device (0): Quadro P600, Compute Capability 6.1
2019-04-18 09:03:21.971218: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2904000000 Hz
2019-04-18 09:03:21.971816: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x55e58577f290 executing computations on platform Host. Devices:
2019-04-18 09:03:21.971842: I tensorflow/compiler/xla/service/service.cc:175]   StreamExecutor device (0): <undefined>, <undefined>
2019-04-18 09:03:21.972102: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1551] Found device 0 with properties:
name: Quadro P600 major: 6 minor: 1 memoryClockRate(GHz): 1.5565
pciBusID: 0000:01:00.0
totalMemory: 1.95GiB freeMemory: 1.91GiB
2019-04-18 09:03:21.972147: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1674] Adding visible gpu devices: 0
2019-04-18 09:03:21.972248: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.0
2019-04-18 09:03:21.973094: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1082] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-04-18 09:03:21.973105: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1088]      0
2019-04-18 09:03:21.973110: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1101] 0:   N
2019-04-18 09:03:21.973279: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1222] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 1735 MB memory) -> physical GPU (device: 0, name: Quadro P600, pci bus id: 0000:01:00.0, compute capability: 6.1)

Any help would be appreciated! I am about to post another question about implementing CUBLAS in R as well because that was giving me slow speed times compared to Intel MKL, but I'm hoping that maybe there's a clear cut reason why even something as well built as TF (compared to hacky R and CUBLAS patching) is being slow with GPU.

Thanks!

EDIT: Following Vlad's suggestion, I wrote up the following script to try and throw some large sized objects and training it, but I think I might not be setting it up correctly because the CPU one in this case even as the size of the matrices are increasing. Any suggestions perhaps?

import time
import warnings
import numpy as np
import scipy

import tensorflow as tf
import tensorflow_probability as tfp

from tensorflow_probability import edward2 as ed
from tensorflow.python.ops import control_flow_ops
from tensorflow_probability import distributions as tfd

np.random.seed(69)

# Handy snippet to reset the global graph and global session.
def reset_g():
    with warnings.catch_warnings():
        warnings.simplefilter('ignore')
        tf.reset_default_graph()
        try:
            sess.close()
        except:
            pass


# Loop over the different number of feature columns
for x_feat in [30, 50, 100, 1000, 10000]:

    y_feat=10;
    # Simulate data
    N = 5000
    inttest = np.ones(N).reshape(N, 1)
    stddev_raw = np.random.uniform(0.01, 0.25, size=y_feat)

    true_int = np.linspace(0.1 ,1., num=y_feat)
    xcols = x_feat
    true_bw = np.random.randn(xcols, y_feat)
    true_X = np.random.randn(N, xcols)
    true_errorcov = np.eye(y_feat)
    np.fill_diagonal(true_errorcov, stddev_raw)

    true_Y = true_int + np.matmul(true_X, true_bw) + \
        np.random.multivariate_normal(mean=np.array([0 for i in range(y_feat)]),
                                      cov=true_errorcov,
                                      size=N)


    ## Our model is:
    ## Y = a + b*X + error where, for N=5000 observations:
    ## Y : 10 outputs;
    ## X : 30,50,100,1000,10000 features
    ## a, b = bias and weights
    ## error: just... error

    # Number of iterations
    N_ITER = 1001

    # Training rate
    lr=0.005

    with tf.device('gpu'):

        # Create data and weights
        model_X = tf.placeholder(dtype=tf.float64, shape=[N, xcols])
        model_Y = tf.placeholder(dtype=tf.float64, shape=[N, y_feat])

        alpha = tf.get_variable(shape=[y_feat], name='alpha', dtype=tf.float64)
        # these two params need shape of one if using trainable distro
        betas = tf.get_variable(shape=[xcols, y_feat], name='beta1', dtype=tf.float64)


        # Yhat
        tf_pred = alpha + tf.matmul(model_X, betas)

        # Make difference of squares (loss fn) [CONVERGES TO TRUTH]
        resid = tf.square(model_Y - tf_pred)
        loss = tf.reduce_sum(resid)

        # Trainer
        opt = tf.train.AdamOptimizer(lr, beta1=0.95, beta2=0.95)
        train = opt.minimize(loss)


    sess = tf.Session()
    sess.run(tf.global_variables_initializer())

    start = time.time()
    for step in range(N_ITER):
        out_l, laws = sess.run([train, loss], feed_dict={model_X: true_X, model_Y: true_Y})

        if step % 500 == 0:
            print('Step: {s}, loss = {l}'.format(
                s=step, l=laws))
    end = time.time()
    print("y_feat: {n}, x_feat: {x2}, Time elapsed: {te}".format(n = y_feat, x2 = x_feat, te = end-start))

    reset_g()

Answer 1

As I said in a comment, the overhead of invoking GPU kernels, and copying data to and from GPU, is very high. For operations on models with very little parameters it is not worth of using GPU since frequency of CPU cores is much higher. If you compare matrix multiplication (this is what DL mostly does), you will see that for large matrices GPU outperforms CPU significantly.

Take a look at this plot. X-axis are the sizes of two square matrices and y-axis is time took to multiply those matrices on GPU and on CPU. As you can see at the beginning, for small matrices the blue line is higher, meaning that it was faster on CPU. But as we increase the size of the matrices the benefit from using GPU increases significantly.

The code to reproduce:

import tensorflow as tf
import time
cpu_times = []
sizes = [1, 10, 100, 500, 1000, 2000, 3000, 4000, 5000, 8000, 10000]
for size in sizes:
    tf.reset_default_graph()
    start = time.time()
    with tf.device('cpu:0'):
        v1 = tf.Variable(tf.random_normal((size, size)))
        v2 = tf.Variable(tf.random_normal((size, size)))
        op = tf.matmul(v1, v2)

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        sess.run(op)
    cpu_times.append(time.time() - start)
    print('cpu time took: {0:.4f}'.format(time.time() - start))

import tensorflow as tf
import time

gpu_times = []
for size in sizes:
    tf.reset_default_graph()
    start = time.time()
    with tf.device('gpu:0'):
        v1 = tf.Variable(tf.random_normal((size, size)))
        v2 = tf.Variable(tf.random_normal((size, size)))
        op = tf.matmul(v1, v2)

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        sess.run(op)
    gpu_times.append(time.time() - start)
    print('gpu time took: {0:.4f}'.format(time.time() - start))

import matplotlib.pyplot as plt

fig, ax = plt.subplots(figsize=(8, 6))
ax.plot(sizes, gpu_times, label='GPU')
ax.plot(sizes, cpu_times, label='CPU')
plt.xlabel('MATRIX SIZE')
plt.ylabel('TIME (sec)')
plt.legend()
plt.show()

Answer 2

Select your Device using tf.device()

with tf.device('/cpu:0'):
    #enter code here of tf data

On a typical system, there are multiple computing devices. In TensorFlow, the supported device types are CPU and GPU. They are represented as strings. For example:

"/cpu:0": The CPU of your machine.
"/device:GPU:0": The GPU of your machine, if you have one.
"/device:GPU:1": The second GPU of your machine, etc.

GPU:

with tf.device('/device:GPU:0'):
  #code here: tf data and model

Reference: Link

Answer 3

to reproduce answer from first comment for tensorflow 2.0+

tf.compat.v1.disable_eager_execution()
cpu_times = []
sizes = [1, 10, 100, 500, 1000, 2000, 3000, 4000, 5000, 8000, 10000]
for size in sizes:
    ops.reset_default_graph()
    start = time.time()
    with tf.device('cpu:0'):
        v1 = tf.Variable(tf.random.normal((size, size)))
        v2 = tf.Variable(tf.random.normal((size, size)))
        op = tf.matmul(v1, v2)

    with tf.compat.v1.Session() as sess:
        sess.run(tf.compat.v1.global_variables_initializer())
        sess.run(op)
    cpu_times.append(time.time() - start)
    print('cpu time took: {0:.4f}'.format(time.time() - start))

import tensorflow as tf
import time

gpu_times = []
for size in sizes:
    ops.reset_default_graph()
    start = time.time()
    with tf.device('gpu:0'):
        v1 = tf.Variable(tf.random.normal((size, size)))
        v2 = tf.Variable(tf.random.normal((size, size)))
        op = tf.matmul(v1, v2)

    with tf.compat.v1.Session() as sess:
        sess.run(tf.compat.v1.global_variables_initializer())
        sess.run(op)
    gpu_times.append(time.time() - start)
    print('gpu time took: {0:.4f}'.format(time.time() - start))

import matplotlib.pyplot as plt

fig, ax = plt.subplots(figsize=(8, 6))
ax.plot(sizes, gpu_times, label='GPU')
ax.plot(sizes, cpu_times, label='CPU')
plt.xlabel('MATRIX SIZE')
plt.ylabel('TIME (sec)')
plt.legend()
plt.show()