CTuner: Automatic Distributed NoSQL Database Tuning with Causal Reinforcement Learning
Configuration tuning is always an effective way to improve the performance of NoSQL DB. Recently, reinforcement learning (RL) has shown great potential in the performance tuning of databases. However, using RL for NoSQL database tuning is still challenging because NoSQL databases usually have a large number of tunable knobs.Moreover, compared with the heuristic algorithm, RL has a cold start problem in the early stage of offline training, which leads to an increase in time cost. On the other hand, it is difficult for RL models to adapt quickly to unseen environments. To address these issues, we propose a NoSQL tuning system named CTuner, which recommends configurations efficiently and effectively. In order to reduce the total cost of tuning, 1) CTuner uses Bayesian Optimization to generate high-quality training samples to solve the cold start problem, and Random Forest is used to select important knobs; 2) CTuner uses causal inference to improve the exploitation strategy of TD3. Meanwhile, a novel Multi-tasking Prioritized Replay Memory which fuses reward and Temporal Difference error is designed; 3) Meta-learning is introduced to improve the adaptability of the tuning model. Our experiments with YCSB benchmark show that CTuner can find a better configuration with the same time cost, with throughput increased by 2.4%-27.4% and 95%-tail latency decreased by 1.2%-13.2%.
| hostname | ip | required | Function | user | password | operating system | Python version |
|---|---|---|---|---|---|---|---|
| CTuner1 | 33.33.33.104 | openjdk-8-jdk | server | ubuntu | ubuntu | Ubuntu18.04 | 3.7 |
| CTuner2 | 33.33.33.179 | openjdk-8-jdk | client | ubuntu | ubuntu | Ubuntu18.04 | 3.7 |
Install openjdk-8-jdk on two hosts
mgt@CTuner1: sudo apt update && sudo apt -y upgrade
mgt@CTuner1: sudo apt-get install openjdk-8-jdk
mgt@CTuner1: java -version
mgt@CTuner2: sudo apt update && sudo apt -y upgrade
mgt@CTuner2: sudo apt-get install openjdk-8-jdk
mgt@CTuner2: java -version##Step2: Install python3.7 for Server Install environment dependencies
mgt@CTuner1: sudo apt-get install zlib1g-dev libbz2-dev libssl-dev libncurses5-dev libsqlite3-dev libreadline-dev tk-dev libgdbm-dev libdb-dev libpcap-dev xz-utils libexpat1-dev liblzma-dev libffi-dev libc6-devCreate an installation directory
mgt@CTuner1: sudo mkdir -p /usr/local/python3.7 Download python3.7 and unzip the download installation package
mgt@CTuner1: wget https://www.python.org/ftp/python/3.7.8/Python-3.7.8.tgz
mgt@CTuner1: tar -zxvf Python-3.7.8.tgzEnter the unzipped directory
mgt@CTuner1: cd Python-3.7.8/Compile and install --prefix: used to specify the installation location --enable-optimizations: for optimization configuration
mgt@CTuner1: ./configure --prefix=/usr/local/python3.7 --enable-optimizationsmgt@CTuner1: makeIn order to prevent replacing the default installed version of the system, use altinstall
mgt@CTuner1: sudo make altinstallCreate soft links to python3.7 and pip3.7
mgt@CTuner1: sudo ln -sf /usr/local/python3.7/bin/python3.7 /usr/bin/python3
mgt@CTuner1: sudo ln -sf /usr/local/python3.7/bin/pip3.7 /usr/bin/pip3Test whether the installation is successful
mgt@CTuner1: python3Install ansible for server
mgt@CTuner1: pip3 install ansible==2.9.27 -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
mgt@CTuner1: sudo apt install ansible
mgt@CTuner1: ansible --versionInstall python for client
mgt@CTuner2: sudo apt-get install pythonAnsible of the server connects to the client through SSH, first generates a public key key on the server node, and then copies it to the client node.
mgt@CTuner1: ssh-keygen
mgt@CTuner1: ls -l /home/ubuntu/.sshThe server copies the public key key to the client node:
mgt@CTuner1: ssh-copy-id [email protected]
mgt@CTuner1: ssh-copy-id [email protected]Test whether it is connected, it is successful without entering a password
mgt@CTuner1: ssh [email protected] Edit /etc/ansible/hosts to add client information
mgt@CTuner1: sudo vim /etc/ansible/hostsAdd content as follows
[nosql]
nosql-tester ansible_ssh_host=33.33.33.104 ansible_ssh_user=ubuntu
nosql-testee ansible_ssh_host=33.33.33.179 ansible_ssh_user=ubuntu
Test that the ping command is executed successfully
mgt@CTuner1: ansible nosql -m pingInstall related dependencies
mgt@CTuner1: sudo apt-get install graphviz graphviz-dev
mgt@CTuner1: pip3 install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.comModify the config file information under the environment
'mongodb': {
'tester': 'mongo-tester', # Need to be modified according to the actual situation
'testee': 'mongo-testee', # Need to be modified according to the actual situation
'host': '33.33.33.179', # Need to be modified according to the actual situation
'user': 'ubuntu', # Need to be modified according to the actual situation
'password': 'ubuntu', # Need to be modified according to the actual situation
'port': 27017,
'database': 'ycsb', # fixed
'memory': 34359738368,
'test_mode': 'synchronization' # fixed
},Take mongodb to obtain high-quality samples as an example, meta_train_01 means calling the configuration file in environment\target\mongodb\tests\meta_train_01.yml
mgt@CTuner1: cd tuner
mgt@CTuner1: python3 meta_preheat.py mongodb meta_train_01Taking mongodb meta training as an example, meta_train_01 means calling the configuration file in environment\target\mongodb\tests\meta_train_01.yml
mgt@CTuner1: cd tuner
mgt@CTuner1: python3 meta_train.py mongodb meta_train_01Taking mongodb evaluating as an example, meta_eval_01 means calling the configuration file in environment\target\mongodb\tests\meta_eval_01.yml
mgt@CTuner1: cd tuner
mgt@CTuner1: python3 meta_evaluate.py mongodb meta_eval_011、 subprocess.CalledProcessError: Command '('lsb_release', '-a')' returned non-zero exit status 1.
Solution:
mgt@CTuner1: sudo find / -name 'lsb_release.py'
mgt@CTuner1: sudo cp /usr/lib/python3/dist-packages/lsb_release.py /usr/local/python3.7/lib/python3.7/