Magpie

This is the code repository for the a big data system parameter automatic optimization paper titled 'Magpie: Efficient Big Data Query System Parameter Optimization based on Pre-selection and Search Pruning Approach'.

Magpie can recommend the best parameter configuration of the big data system (Flink,Spark,etc.)according to the performance target requirements and parameters set by the user and their range of values.

Prerequisites

    CentOS 7.5
    Java 1.8
    Python 3.6.3
    Hadoop 2.6.7
    Hive 2.3.4
    Flink 1.11.0
    Prometheus 2.19.2
    Pushgateway 1.2.0

When installing java, hadoop, hive and Flink, please make sure to set user environment variables for them, such as JAVA_HOME, HADOOP_HOME, FLINK_HOME and PATH

Before the system is running, use Python to load the LightGBM dependency package, install the command: pip install lightgbm

Before the system runs, please make sure that your job can run normally in the Flink cluster

Quick Start

1. Compile and package

   cd Magpie
   mvn clean install -Dmaven.test.skip=true
   

2. System configuration: configure flink parameters and values, inspected performance indicators, performance goals, flink execution jobs and job types and other parameters in conf/config.yaml

   #Flink dir
   flink.dir: /env/flink-1.11.0
   #Flink parameters values
   parameters:
       taskmanager.memory.process.size: [2g,3g,4g,5g,6g,7g,8g,9g,40g,12g,14g,16g,18g,20g,24g,30g]
       taskmanager.numberOfTaskSlots: [2,3,4,5,6,7,8,9,10,11,12,16,20]
       taskmanager.memory.network.fraction: [0.05,0.1,0.15,0.2, 0.25]     
       taskmanager.memory.managed.fraction: [0.2,0.25,0.3,0.35,0.4,0.45,0.5,0.6,0.7]
        parallelism.default: [2,4,8,10,16,20,30,32,40,48,50,60,70,80]
   #performance target
   target: 1.0
   #Flink Job compute model
   flink.job.model: batch
   #job type
   flink.job.type: SQL
   #Flink job submit
   job.submit.cmd: ./bin/flink  run -m yarn-cluster  -c  org.apache.flink.benchmark.Benchmark\  
           ~/target/flink-tpcds-0.1-SNAPSHOT-jar-with-dependencies.jar\    
   		--database tpcds_bin_orc_100\ 
           --queries q7.sql
   

3. Running

   ./bin/start.sh &
   

   After the system is running, you can check whether the Flink job is running normally on Flink Web port 8081 or Yarn port 8088, and you can check job performance data on Prometheus Web port 9091. If you want to stop the system running, execute the command ./bin/stop.sh

4. Operation result: monitor the parameter search process and view the recommended configuration parameter result output

   tail –f logs/task.log （Running）
   tail –f logs/task.out （After running）