Communication Benchmark Models

This module contains simple tensorflow models that can be used for investigating communication costs during a distrbuted tensorflow computation.

Running The Distributed Model

The distributed communication model is implemented in parbuffer.py. To run the model to send a variable between geeker-3 and geeker-4 create run the following:

# To time the commnication of a variable of size 4096 between geeker-3 and geeker-4

# On geeker-4 (parameter server)
$ python parbuffer.py --variable_size=4096 --batch_size=100 --node_name=ps

# On geeker-3
$ python parbuffer.py --variable_size=4096 --batch_size=100 >> results

Running The Single-Machine Model

To test the baseline for the cheap operation on the variable tf.round(y) you can time the code on a single machine using the buffer.py script

$ python buffer.py --variable_size=4096 --batch_size=100

TODO

• combine scripts into one module
• add a parsing script