change pserver to use regular docker and some other tweaks

tools/aws_benchmarking/server/cluster_master.py

@@ -478,6 +478,9 @@ def kickoff_pserver(host, pserver_endpoints_str):
             TASK_NAME=args.task_name,
             COMMAND=args.pserver_command,
             TRAINER_COUNT=args.trainer_count,
+            TRAINER_INDEX=0,
+            # there is no way to use 0.0.0.0:port to start pserver
+            # has to docker --network="host" with host ip to make this work
             SERVER_ENDPOINT=host + ":" + str(args.pserver_port),
             MASTER_ENDPOINT=args.master_server_ip + ":" +
             str(args.master_server_port))
@@ -588,7 +591,7 @@ def start_server(args):
                 logging.info("Received request to return status")
                 with open(args.log_path + "master.log", "r") as logfile:
                     self.wfile.write(logfile.read().strip())
-            elif request_path == "/list_logs":
+            elif request_path == "/list_logs" or request_path == "/logs":
                 self._set_headers()
                 self.wfile.write("\n".join(log_files))
             elif "/log/" in request_path:

tools/aws_benchmarking/server/pserver.sh.template

 #!/bin/bash
-nvidia-docker run -i -p {PSERVER_PORT}:{PSERVER_PORT} -e "SERVER_ENDPOINT={SERVER_ENDPOINT}" -e "MASTER_ENDPOINT={MASTER_ENDPOINT}" -e "TASK_NAME={TASK_NAME}" -e "TRAINING_ROLE=PSERVER" -e "TRAINERS={TRAINER_COUNT}" -e "PSERVER_HOSTS={PSERVER_HOSTS}" -e "PSERVERS={PSERVER_HOSTS}" {DOCKER_IMAGE} {COMMAND}
+docker run --network="host" -i -p {PSERVER_PORT}:{PSERVER_PORT} -e "SERVER_ENDPOINT={SERVER_ENDPOINT}" -e "MASTER_ENDPOINT={MASTER_ENDPOINT}" -e "TASK_NAME={TASK_NAME}" -e "TRAINER_INDEX={TRAINER_INDEX}" -e "TRAINING_ROLE=PSERVER" -e "TRAINER_COUNT={TRAINER_COUNT}" -e "TRAINERS={TRAINER_COUNT}" -e "PSERVER_HOSTS={PSERVER_HOSTS}" -e "PSERVERS={PSERVER_HOSTS}" {DOCKER_IMAGE} {COMMAND} --device CPU
\ No newline at end of file

tools/aws_benchmarking/server/trainer.sh.template

 #!/bin/bash 
-nvidia-docker run -i  -e "MASTER_ENDPOINT={MASTER_ENDPOINT}" -e "TASK_NAME={TASK_NAME}" -e "TRAINER_COUNT={TRAINER_COUNT}" -e "TRAINER_INDEX={TRAINER_INDEX}" -e "TRAINING_ROLE=TRAINER"  -e "PSERVER_HOSTS={PSERVER_HOSTS}" {DOCKER_IMAGE} {COMMAND}
+nvidia-docker run -i  -e "MASTER_ENDPOINT={MASTER_ENDPOINT}" -e "TASK_NAME={TASK_NAME}" -e "TRAINER_COUNT={TRAINER_COUNT}" -e "TRAINERS={TRAINER_COUNT}" -e "TRAINER_INDEX={TRAINER_INDEX}" -e "TRAINING_ROLE=TRAINER"  -e "PSERVER_HOSTS={PSERVER_HOSTS}" {DOCKER_IMAGE} {COMMAND} --device GPU
\ No newline at end of file