Intro

Code accompanying the HPC CHINA 2018 paper: "Performance Optimization of Lustre File System Based on Reinforcement Learning". Based on data access characteristics of high energy physics calculation, we propose an automated parameter tuning method using the reinforcement learning.

Dependencies

Python:

• Pytorch 0.4
• Zeromq
• Numpy
• Matplotlib

IOzone

How to Use

The storage_drl directory should be placed on the target nodes, and storage_drl_agents directory should be placed on the policy node.

On the policy node, use drl.sh to start agent, you can choose agents here. For example,

./drl.sh start ppo

Stop agent:

./drl.sh stop

Observe agent's state:

./drl.sh status

On the target nodes, use lustre.sh to start monitor and actor. For example,

./lustre.sh start

Others ibid.

Test

You should use IOzone to initiate cluster testing, for this you need to configure password-free login between nodes. Once you configure these, the drl_test directory should be placed on the test initiator node, then use

./iozone.sh start

Others ibid.

Log

The running process will be recorded in the log directory, you can observe the log file for debugging. The log format is as follows:

[ 2018-06-19 19:35:35,192 - ppo ]  -  received {'states': [0.0, 50364416.0, 0.0, 0.0, 0.0, 0.0, 3051.0, 22175142.0, 22175142.0, 0.0, 39317504.0, 0.0, 0.0, 0.0, 0.0, 2468.0, 4915496.0, 4915496.0, 0.0, 70311936.0, 0.0, 0.0, 0.0, 0.0, 2721.0, 8694652.0, 8694652.0, 0.0, 86302720.0, 0.0, 0.0, 0.0, 0.0, 998.0, 18650515.0, 18650515.0, 0.0, 41377792.0, 0.0, 0.0, 0.0, 0.0, 3532.0, 9820971.0, 9820971.0, 0.0, 40869888.0, 0.0, 0.0, 0.0, 0.0, 7159.0, 4780955.0, 4780955.0, 0.0, 79081472.0, 0.0, 0.0, 0.0, 0.0, 901.0, 25180824.0, 25180824.0, 0.0, 82411520.0, 0.0, 0.0, 0.0, 0.0, 2199.0, 40397523.0, 40397523.0, 0.0, 75210752.0, 0.0, 0.0, 0.0, 0.0, 6964.0, 1817595.0, 1817595.0, 0.0, 80928768.0, 0.0, 0.0, 0.0, 0.0, 1495.0, 527019.0, 527019.0, 0.0, 74575872.0, 0.0, 0.0, 0.0, 0.0, 9865.0, 3865213.0, 3865213.0, 0.0, 49229824.0, 0.0, 0.0, 0.0, 0.0, 2228.0, 17157236.0, 17157236.0, 0.0, 78893056.0, 0.0, 0.0, 0.0, 0.0, 4611.0, 5746131.0, 5746131.0, 0.0, 41152512.0, 0.0, 0.0, 0.0, 0.0, 3320.0, 8740776.0, 8740776.0, 0.0, 82092032.0, 0.0, 0.0, 0.0, 0.0, 9060.0, 3816115.0, 3816115.0, 0.0, 38621184.0, 0.0, 0.0, 0.0, 0.0, 1973.0, 7240485.0, 7240485.0, 0.0, 38486016.0, 0.0, 0.0, 0.0, 0.0, 33353.0, 11697863.0, 11697863.0, 0.0, 77275136.0, 0.0, 0.0, 0.0, 0.0, 1120.0, 8042394.0, 8042394.0, 0.0, 72835072.0, 0.0, 0.0, 0.0, 0.0, 6843.0, 6022215.0, 6022215.0, 0.0, 35414016.0, 0.0, 0.0, 0.0, 0.0, 3710.0, 7993143.0, 7993143.0, 0.0, 36880384.0, 0.0, 0.0, 0.0, 0.0, 13861.0, 1998617.0, 1998617.0, 0.0, 80691200.0, 0.0, 0.0, 0.0, 0.0, 1247.0, 2846486.0, 2846486.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1280.0, 1280.0, 2077.0], 'reward': 0.0, 'name': 'lustre_lustretest1'} from lustre_lustretest1
[ 2018-06-19 19:35:35,201 - ppo ]  -  send 5 to lustre_lustretest1