上次做了Hadoop集群的性能测试,因为主要的大数据开发工作在Spark上,这次做一下Spark的性能测试。
代码参考:spark-terasort
因为使用的CDH6.0.1,Spark版本2.2.0,代码需要做一些修改,这里已经在Spark2.3源码下修改好并编译打包,放到了Spark的examples里,可以替换Spark的examples。
链接:网盘下载
参数的设置
参考:Spark资源优化
目录
1. TeraGen生成随机数
生成随机数,Teragen指定数量参数为 k,m,g,t,如100g。
将结果输出到HDFS /benchmarks/spark-test/terasort-input。
sudo -uhdfs spark-submit \
--class org.apache.spark.examples.TeraGen \
--master yarn \
--deploy-mode cluster \
--driver-memory 2g \
--num-executors 6 \
--executor-cores 5 \
--executor-memory 5g \
--conf spark.default.parallelism=240 \
--conf spark.serializer=org.apache.spark.serializer.KryoSerializer \
/opt/cloudera/parcels/CDH/lib/spark/examples/jars/spark-examples_2.11-2.3.3.jar \
500g /benchmarks/spark-test/terasort-input
生成结果:500GB数据1小时,平均8530m/分,142m/秒。
优化后的提交
sudo -uhdfs spark-submit \
--class org.apache.spark.examples.TeraGen \
--master yarn \
--deploy-mode cluster \
--driver-memory 2g \
--driver-cores 5 \
--num-executors 20 \
--executor-cores 4 \
--executor-memory 2560m \
--conf spark.default.parallelism=160 \
--conf spark.shuffle.file.buffer=128k \
--conf spark.serializer=org.apache.spark.serializer.KryoSerializer \
/opt/cloudera/parcels/CDH/lib/spark/examples/jars/spark-examples_2.11-2.3.3.jar \
100g /benchmarks/spark-test/terasort-input
100GB数据12分钟
2. TeraSort排序
sudo -uhdfs spark-submit \
--class org.apache.spark.examples.TeraSort \
--master yarn \
--deploy-mode cluster \
--driver-memory 2g \
--num-executors 20 \
--executor-cores 4 \
--executor-memory 2g \
--conf spark.default.parallelism=160 \
--conf spark.serializer=org.apache.spark.serializer.KryoSerializer \
/opt/cloudera/parcels/CDH/lib/spark/examples/jars/spark-examples_2.11-2.3.3.jar \
/benchmarks/spark-test/terasort-input /benchmarks/spark-test/terasort-output
排序结果:500G 1小时33分。
100G16分钟
sudo -uhdfs spark-submit \
--class org.apache.spark.examples.TeraSort \
--master yarn \
--deploy-mode cluster \
--driver-memory 2g \
--driver-cores 5 \
--num-executors 20 \
--executor-cores 4 \
--executor-memory 2560m \
--conf spark.default.parallelism=240 \
--conf spark.shuffle.file.buffer=128k \
--conf spark.executor.extraJavaOptions="-XX:+UseG1GC -XX:MaxGCPauseMillis=100 -XX:-ResizePLAB -XX:+ParallelRefProcEnabled -XX:ParallelGCThreads=20 -XX:ConcGCThreads=10" \
--conf spark.serializer=org.apache.spark.serializer.KryoSerializer \
/opt/cloudera/parcels/CDH/lib/spark/examples/jars/spark-examples_2.11-2.3.3.jar \
/benchmarks/spark-test/terasort-input /benchmarks/spark-test/terasort-output
3.TeraValidate验证
sudo -uhdfs spark-submit \
--class org.apache.spark.examples.TeraValidate \
--master yarn \
--deploy-mode cluster \
--driver-memory 2g \
--driver-cores 5 \
--num-executors 20 \
--executor-cores 4 \
--executor-memory 2560m \
--conf spark.default.parallelism=240 \
--conf spark.serializer=org.apache.spark.serializer.KryoSerializer \
/opt/cloudera/parcels/CDH/lib/spark/examples/jars/spark-examples_2.11-2.3.3.jar \
/benchmarks/spark-test/terasort-output
验证500G数据10分钟左右。
client模式结果会输出在控制台,cluster模式结果输出在YARN的logs里。
num records: 4814814810
checksum: 8f7e4a4b7086a04c
partitions are properly sorted
100G 3分钟
num records: 1000000000
checksum: 1dcd3a102148ad22