Tensorflow: simultaneous prediction on GPU and CPU

Question

I'm working with tensorflow and I want to speed up the prediction phase of a pre-trained Keras model (I'm not interested in the training phase) by using simultaneously the CPU and one GPU.

I tried to create 2 different threads that feed two different tensorflow sessions (one that runs on CPU and the other that runs on GPU). Each thread feeds a fixed number of batches (eg if we have an overall of 100 batches, I want to assign 20 batches for CPU and 80 on GPU, or any possible combination of the two) in a loop and combine the result. It would be better if the split was done automatically.

However even in this scenario, it seems that the batches are being fed in a synchronous way, because even sending few batches to the CPU and computing all the others in the GPU (with the GPU as bottleneck) I observed that the overall prediction time is always higher with respect to the test made only using the GPU.

I would expect it to be faster because when only the GPU is working the CPU usage is about 20-30%, thus there is some CPU available to speed up the computation.

I read a lot of discussions but they all deal with parallelism with multiple GPUs and not between GPU and CPU.

Here is a sample of the code I have written: the tensor_cpu and tensor_gpu objects are loaded from the same Keras model in this way:

with tf.device('/gpu:0'):
    model_gpu = load_model('model1.h5')
    tensor_gpu = model_gpu(x)

with tf.device('/cpu:0'):
    model_cpu = load_model('model1.h5')
    tensor_cpu = model_cpu(x) 

Then the prediction is done as following:

def predict_on_device(session, predict_tensor, batches):
    for batch in batches:
        session.run(predict_tensor, feed_dict={x: batch})


def split_cpu_gpu(batches, num_batches_cpu, tensor_cpu, tensor_gpu):
    session1 = tf.Session(config=tf.ConfigProto(log_device_placement=True))
    session1.run(tf.global_variables_initializer())
    session2 = tf.Session(config=tf.ConfigProto(log_device_placement=True))
    session2.run(tf.global_variables_initializer())

    coord = tf.train.Coordinator()

    t_cpu = Thread(target=predict_on_device, args=(session1, tensor_cpu, batches[:num_batches_cpu]))
    t_gpu = Thread(target=predict_on_device, args=(session2, tensor_gpu, batches[num_batches_cpu:]))

    t_cpu.start()
    t_gpu.start()

    coord.join([t_cpu, t_gpu])

    session1.close()
    session2.close()

How can I achieve this CPU/GPU parallelization? I think I'm missing something.

Any kind of help would be very appreciated!

Answer 1

Here's my code that demonstrates how CPU and GPU execution can be done in parallel:

import tensorflow as tf
import numpy as np
from time import time
from threading import Thread

n = 1024 * 8

data_cpu = np.random.uniform(size=[n//16, n]).astype(np.float32)
data_gpu = np.random.uniform(size=[n    , n]).astype(np.float32)

with tf.device('/cpu:0'):
    x = tf.placeholder(name='x', dtype=tf.float32)

def get_var(name):
    return tf.get_variable(name, shape=[n, n])

def op(name):
    w = get_var(name)
    y = x
    for _ in range(8):
        y = tf.matmul(y, w)
    return y

with tf.device('/cpu:0'):
    cpu = op('w_cpu')

with tf.device('/gpu:0'):
    gpu = op('w_gpu')

def f(session, y, data):
    return session.run(y, feed_dict={x : data})


with tf.Session(config=tf.ConfigProto(log_device_placement=True, intra_op_parallelism_threads=8)) as sess:
    sess.run(tf.global_variables_initializer())

    coord = tf.train.Coordinator()

    threads = []

    # comment out 0 or 1 of the following 2 lines:
    threads += [Thread(target=f, args=(sess, cpu, data_cpu))]
    threads += [Thread(target=f, args=(sess, gpu, data_gpu))]

    t0 = time()

    for t in threads:
        t.start()

    coord.join(threads)

    t1 = time()


print t1 - t0

The timing results are:

• CPU thread: 4-5s (will vary by machine, of course).

• GPU thread: 5s (It does 16x as much work).

• Both at the same time: 5s

Note that there was no need to have 2 sessions (but that worked for me too).

The reasons you might be seeing different results could be

• some contention for system resources (GPU execution does consume some host system resources, and if running the CPU thread crowds it, that could worsen the performance)

• incorrect timing

• part of your model can only run on GPU/CPU

• bottleneck elsewhere

• some other problem