samjabrahams/tensorflow-white-paper-notes

Validation: consider memory in placement decision

larissa95 opened this issue · 2 comments

"Final selection of devices is determined using similar heuristics as described in Section 3.2, ensuring fast execution while taking device restrictions, such as memory, into account."

Where have you found that in the code? (That they consider the memory)

This is just a description of the algorithm described in the white paper: the device placement algorithm isn't available in the open-sourced implementation of TensorFlow.

Mhmm, do you know how a device is chosen for the nodes in the simple placer (where and how do they prioritize the devices -> it is probably in the union-find algorithm?):
status = colocation_graph.GetDevicesForNode(node, &devices);
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/common_runtime/simple_placer.cc#L744