Need a baseline performance benchmark for distributed training.

[typhoonzero]

We need a baseline for benchmarking distributed training, the baseline is:

• Select a model (e.g. vgg16)
• Data Parallelism, sync SGD
• Same cluster hardware

the variables to compare are:

• Different frameworks (PaddlePaddle vs. others)
• The number of nodes
• batch size

The performance baseline can be used to judge our framework's distributed training performance than. Both v1 and fluid should be tested.

[typhoonzero]

Closing due to #7539 is complete.