Profiler traces are not the expected duration

[yuzu-ido]

Hi all, when running profiling on a jax training script with jax.profiler.start_server() and

from tensorflow.python.profiler import profiler_client 
options = profiler_v2.ProfilerOptions(
       host_tracer_level=2,
       device_tracer_level=1,
       python_tracer_level=1,
   )
profiler_client.trace(
       address,
       duration_ms=duration_ms,
       logdir=str(_LOG_DIR),
       options=options,
   )

with a duration such as 30000 ms (30 secs), the resulting file after converting to json is variable and generally not the intended duration. Often it is shorter, such as 7 seconds, but on occasion it is longer, such as 40. The trace call does wait for completion for the desired time. Any ideas how this is caused- by the server, client, etc.? Thank you!