How to read trace file(timeline) in Tensorflow

Question

The timeline trace file is introduced here: https://www.tensorflow.org/versions/r1.1/performance/xla/jit

It seems useful for performance analyzing. But there is something I can't understand.

1, What does "pid" stand for while there are:

    /device:GPU:0/stream:35 Compute (pid 3)
    /device:GPU:0/stream:all Compute (pid 9)
    /job:localhost/replica:0/task:0/device:CPU:0 Compute (pid 5)

2, A same OP can appear in different pids with absolutely different timestamp. What does that mean?

Can anybody help me? Or is there any document about this? Thanks very much!

Answer 1

1. 'pid' stands for process identifier. According to a comment from a developer here, "All of the numeric 'PID's and 'TID's in the UI should be ignored- they were just invented to get CTV [Chrome Trace Visualizer] to lay out the trace better but there's no way to stop the GUI displaying them!"

2. This probably means that the OP is running multiple times.