-
You can add your benchmark file under tensorflow/benchmarks/models directory. The benchmark should accept
task_id
,job_name
,ps_hosts
andworker_hosts
flags. You can copy-paste the following flag definitions:tf.app.flags.DEFINE_integer("task_id", None, "Task index, should be >= 0.") tf.app.flags.DEFINE_string("job_name", None, "job name: worker or ps") tf.app.flags.DEFINE_string("ps_hosts", None, "Comma-separated list of hostname:port pairs") tf.app.flags.DEFINE_string("worker_hosts", None, "Comma-separated list of hostname:port pairs")
-
Report benchmark values by calling
store_data_in_json
from your benchmark code. This function is defined in benchmark_util.py -
Create a Dockerfile that sets up dependencies and runs your benchmark. For example, see Dockerfile.alexnet_distributed_test
-
Add the benchmark to benchmark_configs.yml
- Set
benchmark_name
to a descriptive name for your benchmark and make sure it is unique. - Set
worker_count
andps_count
. - Set
docker_file
to the Dockerfile path starting withbenchmarks/
directory. - Optionally, you can pass flags to your benchmark by adding
args
list.
- Set
-
Send PR with the changes to annarev.
For any questions, please contact annarev@google.com.