在cifar10_estimator代码上测试，多个GPU（1080Ti）不会加快张量流中的训练

Question

我试图在2或3 1080Ti上测试cifar10_estimator的多GPU版本的性能，但没有提高速度。

我找到了一些有关硬件here的有用信息，但仍然困惑如何解决。

我的环境：

• Ubuntu版本= 16.04.5 LTS（Xenial Xerus）
• Python3
• CUDA_VERSION = 9.0.176
• tensorflow-gpu = 1.11.0

GPU信息：

nvidia-smi topo -m

    GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    CPU Affinity
GPU0     X  PIX PHB PHB SYS SYS SYS SYS 0-7
GPU1    PIX  X  PHB PHB SYS SYS SYS SYS 0-7
GPU2    PHB PHB  X  PIX SYS SYS SYS SYS 0-7
GPU3    PHB PHB PIX  X  SYS SYS SYS SYS 0-7
GPU4    SYS SYS SYS SYS  X  PIX PHB PHB 8-15
GPU5    SYS SYS SYS SYS PIX  X  PHB PHB 8-15
GPU6    SYS SYS SYS SYS PHB PHB  X  PIX 8-15
GPU7    SYS SYS SYS SYS PHB PHB PIX  X  8-15

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe switches (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing a single PCIe switch
  NV#  = Connection traversing a bonded set of # NVLinks

1 gpu bach_size = 128

INFO:tensorflow:loss = 2.2576141, step = 200 (3.729 sec)
INFO:tensorflow:learning_rate = 0.1, loss = 2.2576141 (3.729 sec)
INFO:tensorflow:Average examples/sec: 2821.06 (2858.65), step = 200
INFO:tensorflow:Average examples/sec: 2847.23 (3496.06), step = 210
INFO:tensorflow:Average examples/sec: 2857.91 (3102.29), step = 220
INFO:tensorflow:Average examples/sec: 2867.04 (3083.62), step = 230
INFO:tensorflow:Average examples/sec: 2889.21 (3514.15), step = 240
INFO:tensorflow:Average examples/sec: 2913.15 (3636.28), step = 250
INFO:tensorflow:Average examples/sec: 2915.99 (2988.94), step = 260
INFO:tensorflow:Average examples/sec: 2901.94 (2578.95), step = 270
INFO:tensorflow:Average examples/sec: 2888.87 (2575.46), step = 280
INFO:tensorflow:Average examples/sec: 2892.13 (2986.66), step = 290
INFO:tensorflow:global_step/sec: 24.25

2 gpu bach_size = 256

INFO:tensorflow:loss = 2.4630964, step = 200 (5.971 sec)
INFO:tensorflow:learning_rate = 0.1, loss = 2.4630964 (5.971 sec)
INFO:tensorflow:Average examples/sec: 3255.68 (4296.71), step = 200
INFO:tensorflow:Average examples/sec: 3297.51 (4437.93), step = 210
INFO:tensorflow:Average examples/sec: 3332.15 (4275.33), step = 220
INFO:tensorflow:Average examples/sec: 3363.86 (4254.65), step = 230
INFO:tensorflow:Average examples/sec: 3395.09 (4316.94), step = 240
INFO:tensorflow:Average examples/sec: 3418.44 (4094.23), step = 250
INFO:tensorflow:Average examples/sec: 3447.17 (4364.24), step = 260
INFO:tensorflow:Average examples/sec: 3474.56 (4379.02), step = 270
INFO:tensorflow:Average examples/sec: 3492.73 (4067.13), step = 280
INFO:tensorflow:Average examples/sec: 3514.19 (4244.23), step = 290
INFO:tensorflow:global_step/sec: 16.6026

3 gpu bach_size = 384

INFO:tensorflow:loss = 2.0980535, step = 200 (9.329 sec)
INFO:tensorflow:learning_rate = 0.1, loss = 2.0980535 (9.329 sec)
INFO:tensorflow:Average examples/sec: 3214.65 (4165.7), step = 200
INFO:tensorflow:Average examples/sec: 3272.85 (5130.99), step = 210
INFO:tensorflow:Average examples/sec: 3324.15 (4955.13), step = 220
INFO:tensorflow:Average examples/sec: 3376.65 (5174.76), step = 230
INFO:tensorflow:Average examples/sec: 3425.48 (5132.15), step = 240
INFO:tensorflow:Average examples/sec: 3468.29 (4954.35), step = 250
INFO:tensorflow:Average examples/sec: 3509.91 (5014.23), step = 260
INFO:tensorflow:Average examples/sec: 3544.29 (4755.56), step = 270
INFO:tensorflow:Average examples/sec: 3579.69 (4901.39), step = 280
INFO:tensorflow:Average examples/sec: 3617.84 (5156.66), step = 290
INFO:tensorflow:global_step/sec: 13.1009

Answer 1

我想我现在可以回答我的问题。如果我想要多个GPU更高的性能，我应该签出https://github.com/tensorflow/benchmarks/。有关我在this issue上的测试结果，请参考tf_cnn_benchmarks。