Specify gpu in Tensorflow code: /gpu:0 is always working?

Question

I have 3 graphics cards in my workstation, one of them is Quadro K620, and the other two are Titan X. Now I would like to run my tensorflow code in one of the graphics card, so that I can leave the others idle for another task.

However, regardless of setting tf.device('/gpu:0') or tf.device('/gpu:1'), I found the 1st Titan X graphics card is always working, I don't know why.

import argparse
import os
import time
import tensorflow as tf
import numpy as np
import cv2

from Dataset import Dataset
from Net import Net

FLAGS = None

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('--foldername', type=str, default='./data-large/')
    parser.add_argument('--batch_size', type=int, default=100)
    parser.add_argument('--num_epoches', type=int, default=100)
    parser.add_argument('--learning_rate', type=float, default=0.5)

    FLAGS = parser.parse_args()
    net = Net(FLAGS.batch_size, FLAGS.learning_rate)

    with tf.Graph().as_default():
        # Dataset is a class for encapsulate the input pipeline
        dataset = Dataset(foldername=FLAGS.foldername,
                              batch_size=FLAGS.batch_size,
                              num_epoches=FLAGS.num_epoches)

        images, labels = dataset.samples_train

        ## The following code defines the network and train
        with tf.device('/gpu:0'): # <==== THIS LINE
            logits = net.inference(images)
            loss = net.loss(logits, labels)
            train_op = net.training(loss)

            init_op = tf.group(tf.initialize_all_variables(), tf.initialize_local_variables())
            sess = tf.Session()
            sess.run(init_op)
            coord = tf.train.Coordinator()
            threads = tf.train.start_queue_runners(sess=sess, coord=coord)
            start_time = time.time()
            try:
                step = 0
                while not coord.should_stop():
                    _, loss_value = sess.run([train_op, loss])
                    step = step + 1
                    if step % 100 == 0:
                        format_str = ('step %d, loss = %.2f, time: %.2f seconds')
                        print(format_str % (step, loss_value, (time.time() - start_time)))
                        start_time = time.time()
            except tf.errors.OutOfRangeError:
                print('done')
            finally:
                coord.request_stop()

            coord.join(threads)
            sess.close()

Regarding to the line "<=== THIS LINE:"

If I set tf.device('/gpu:0'), the monitor says:

|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro K620         Off  | 0000:03:00.0      On |                  N/A |
| 34%   45C    P0     2W /  30W |    404MiB /  1993MiB |      5%      Default |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GTX TIT...  Off  | 0000:04:00.0     Off |                  N/A |
| 22%   39C    P2   100W / 250W |  11691MiB / 12206MiB |      8%      Default |
+-------------------------------+----------------------+----------------------+
|   2  GeForce GTX TIT...  Off  | 0000:81:00.0     Off |                  N/A |
| 22%   43C    P2    71W / 250W |    111MiB / 12206MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

showing the 1st Titan X card is working.

If I set tf.device('/gpu:1'), the monitor says:

|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro K620         Off  | 0000:03:00.0      On |                  N/A |
| 34%   45C    P0     2W /  30W |    411MiB /  1993MiB |      3%      Default |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GTX TIT...  Off  | 0000:04:00.0     Off |                  N/A |
| 22%   52C    P2    73W / 250W |  11628MiB / 12206MiB |     12%      Default |
+-------------------------------+----------------------+----------------------+
|   2  GeForce GTX TIT...  Off  | 0000:81:00.0     Off |                  N/A |
| 22%   42C    P2    71W / 250W |  11628MiB / 12206MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

showing that the two Titan X cards are working, not the 2nd Titan X alone.

So any reason behind this and how to specify the gpu I want my program to run in?

Answer 1

Just a guess, but the default behavior for a tf.train.Optimizer object (which I expect is created in net.training(loss)) when you call minimize() is colocate_gradients_with_ops=False. This may lead to the backpropagation ops being placed on the default device, which will be /gpu:0.

To work out if this is happening, you can iterate over sess.graph_def and look for nodes that either have /gpu:0 in the NodeDef.device field, or have an empty device field (in which case they will be placed on /gpu:0 by default).

Another option for checking what devices are being used is to use the output_partition_graphs=True option when running your step. This shows what devices TensorFlow is actually using (instead of, in sess.graph_def, what devices your program is requesting), and should show exactly what nodes are running on /gpu:0.