UniTune

This is the source code to the paper "A Unified and Efficient Coordinating Framework for Autonomous DBMS Tuning". Please refer to the paper for the experimental details.

Table of Content

Benchmark Preparation
Installation
Experimental Evaluation
Extendability

Installation

Preparations: Python == 3.7
Install packages
```
pip install -r requirements.txt
```

Benchmark Preparation

Join-Order-Benchmark (JOB)

Download IMDB Data Set from http://homepages.cwi.nl/~boncz/job/imdb.tgz.

Follow the instructions to load data

For MySQL: https://github.com/winkyao/join-order-benchmark.

TPC-H

Download TPC-H Data Set from https://www.tpc.org/.

Follow the instructions to load data:

For MySQL: https://github.com/catarinaribeir0/queries-tpch-dbgen-mysql

Experimental Evaluation

Setup

Please modify the user-specified parameters in config.ini before the experiments.

To specify the database connection information, please modify the following parameters:

[database]
dbtype = mysql
host = 127.0.0.1
port = 3306
user = root
passwd =
dbname = tpch
sock = /mysql/base/mysql.sock
cnf = /MultiTune/default.cnf
# mysql related
mysqld = mysql/mysqlInstall/bin/mysqld
# knob related
knob_config_file =  /MultiTune/knob_configs/mysql_new.json
knob_num = 50
# workload name in ['TPCH', 'JOB']
workload_name = TPCH
# workload execution time constraint in sec
workload_timeout = 600
# workload queries list
workload_qlist_file = /MultiTune/scripts/tpch_queries_list_0.txt
# workload queries directory
workload_qdir = /MultiTune/queries/tpch_queries_mysql_0/
# workload run_scripts directory
scripts_dir = /MultiTune/scripts/
# view generation dir
q_mv_file = /MultiTune/advisor/av_files/trainset/q_mv_list.txt
mv_trainset_dir = /MultiTune/advisor/av_files/trainset

To specify the tuning setting ,please modify the following parameters:

[tuning]
task_id = tpch_test
components = {'knob': 'OtterTune', 'index':'DBA-Bandit', 'query':'LearnedRewrite'}
tuning_budget = 108000
sub_budget = 1200
context = True
context_type = im
context_pca_components = 5
output_file = optimize_history/tpch_test.res
index_budget = 6500
arm_method = ts
ts_use_window = True
window_size = 7
cost_aware = True
max_runs = 200
block_runs = 2

init_runs = 10
converage_judge = False
test = False

Run

To start a training session, please run:

python main.py

Extendability

To add a new agent, a developer needs inherit the corresponding base class and overrides its functions. Please refer to MultiTune/advisor/adviser_example.py for an example.

ZhengtongYan / UniTune