../_images/kyuubi_logo.png

3. Z-order Benchmark

Z-order is a technique that allows you to map multidimensional data to a single dimension. We did a performance test.

For this test ,we used aliyun Databricks Delta test case https://help.aliyun.com/document_detail/168137.html?spm=a2c4g.11186623.6.563.10d758ccclYtVb.

Prepare data for the three scenarios:

  1. 10 billion data and 2 hundred files (parquet files): for big file(1G)

  2. 10 billion data and 1 thousand files (parquet files): for medium file(200m)

  3. 1 billion data and 10 thousand files (parquet files): for smaller file(200k)

Test env: spark-3.1.2 hadoop-2.7.2 kyuubi-1.4.0

Test step:

Step1: create hive tables.

spark.sql(s"drop database if exists $dbName cascade")
spark.sql(s"create database if not exists $dbName")
spark.sql(s"use $dbName")
spark.sql(s"create table $connRandomParquet (src_ip string, src_port int, dst_ip string, dst_port int) stored as parquet")
spark.sql(s"create table $connOrderbyOnlyIp (src_ip string, src_port int, dst_ip string, dst_port int) stored as parquet")
spark.sql(s"create table $connOrderby (src_ip string, src_port int, dst_ip string, dst_port int) stored as parquet")
spark.sql(s"create table $connZorderOnlyIp (src_ip string, src_port int, dst_ip string, dst_port int) stored as parquet")
spark.sql(s"create table $connZorder (src_ip string, src_port int, dst_ip string, dst_port int) stored as parquet")
spark.sql(s"show tables").show(false)

Step2: prepare data for parquet table with three scenarios, we use the following code.

def randomIPv4(r: Random) = Seq.fill(4)(r.nextInt(256)).mkString(".")
def randomPort(r: Random) = r.nextInt(65536)

def randomConnRecord(r: Random) = ConnRecord(
  src_ip = randomIPv4(r), src_port = randomPort(r),
  dst_ip = randomIPv4(r), dst_port = randomPort(r))

Step3: do optimize with z-order only ip and do optimize with order by only ip, sort column: src_ip, dst_ip and shuffle partition just as file numbers.

INSERT overwrite table conn_order_only_ip select src_ip, src_port, dst_ip, dst_port from conn_random_parquet order by src_ip, dst_ip;
OPTIMIZE conn_zorder_only_ip ZORDER BY src_ip, dst_ip;

Step4: do optimize with z-order and do optimize with order by, sort column: src_ip, src_port, dst_ip, dst_port and shuffle partition just as file numbers.

INSERT overwrite table conn_order select src_ip, src_port, dst_ip, dst_port from conn_random_parquet order by src_ip, src_port, dst_ip, dst_port;
OPTIMIZE conn_zorder ZORDER BY src_ip, src_port, dst_ip, dst_port;

The complete code is as follows:

./spark-shell
import org.apache.spark.SparkConf
import org.apache.spark.sql.SparkSession

case class ConnRecord(src_ip: String, src_port: Int, dst_ip: String, dst_port: Int)

val  conf  = new SparkConf().setAppName("zorder_test")
val spark = SparkSession.builder().config(conf).enableHiveSupport().getOrCreate()
import spark.implicits._

val sc = spark.sparkContext
sc.setLogLevel("WARN")
//ten billion rows and two hundred files
val numRecords = 10*1000*1000*1000L
val numFiles = 200

val dbName = s"zorder_test_$numFiles"
val baseLocation = s"hdfs://localhost:9000/zorder_test/$dbName/"
val connRandomParquet = "conn_random_parquet"
val connZorderOnlyIp = "conn_zorder_only_ip"
val connZorder = "conn_zorder"
spark.conf.set("spark.sql.shuffle.partitions", numFiles)
spark.conf.get("spark.sql.shuffle.partitions")
spark.conf.set("spark.sql.hive.convertMetastoreParquet",false)
spark.sql(s"drop database if exists $dbName cascade")
spark.sql(s"create database if not exists $dbName")
spark.sql(s"use $dbName")
spark.sql(s"create table $connRandomParquet (src_ip string, src_port int, dst_ip string, dst_port int) stored as parquet")
spark.sql(s"create table $connOrderbyOnlyIp (src_ip string, src_port int, dst_ip string, dst_port int) stored as parquet")
spark.sql(s"create table $connOrderby (src_ip string, src_port int, dst_ip string, dst_port int) stored as parquet")
spark.sql(s"create table $connZorderOnlyIp (src_ip string, src_port int, dst_ip string, dst_port int) stored as parquet")
spark.sql(s"create table $connZorder (src_ip string, src_port int, dst_ip string, dst_port int) stored as parquet")
spark.sql(s"show tables").show(false)

import scala.util.Random
// Function for preparing Zorder_Test data
def randomIPv4(r: Random) = Seq.fill(4)(r.nextInt(256)).mkString(".")
def randomPort(r: Random) = r.nextInt(65536)

def randomConnRecord(r: Random) = ConnRecord(
src_ip = randomIPv4(r), src_port = randomPort(r),
dst_ip = randomIPv4(r), dst_port = randomPort(r))

val df = spark.range(0, numFiles, 1, numFiles).mapPartitions { it =>
val partitionID = it.toStream.head
val r = new Random(seed = partitionID)
Iterator.fill((numRecords / numFiles).toInt)(randomConnRecord(r))
}

df.write
.mode("overwrite")
.format("parquet")
.insertInto(connRandomParquet)

spark.read.table(connRandomParquet)
.write
.mode("overwrite")
.format("parquet")
.insertInto(connZorderOnlyIp)

spark.read.table(connRandomParquet)
.write
.mode("overwrite")
.format("parquet")
.insertInto(connZorder)
spark.stop()

Z-order Optimize statement:

set spark.sql.hive.convertMetastoreParquet=false;

OPTIMIZE conn_zorder_only_ip ZORDER BY src_ip, dst_ip;

OPTIMIZE zorder_test.conn_zorder ZORDER BY src_ip, src_port, dst_ip, dst_port;

ORDER BY statement:

INSERT overwrite table conn_order_only_ip select src_ip, src_port, dst_ip, dst_port from conn_random_parquet order by src_ip, dst_ip;

INSERT overwrite table conn_order select src_ip, src_port, dst_ip, dst_port from conn_random_parquet order by src_ip, src_port, dst_ip, dst_port;

Query statement:

set spark.sql.hive.convertMetastoreParquet=true;

select count(*) from conn_random_parquet where src_ip like '157%' and dst_ip like '216.%';

select count(*) from conn_zorder_only_ip where src_ip like '157%' and dst_ip like '216.%';

select count(*) from conn_zorder where src_ip like '157%' and dst_ip like '216.%';

3.1. Benchmark result

We have done two performance tests: one is to compare the efficiency of Z-order Optimize and Order by Sort, and the other is to query based on the optimized Z-order by data and Random data.

3.1.1. Efficiency of Z-order Optimize and Order-by Sort

10 billion data and 1000 files and Query resource: 200 core 600G memory

Z-order by or order by only ip:

Table row count optimize time
conn_order_only_ip 10,000,000,000 1591.99 s
conn_zorder_only_ip 10,000,000,000 8371.405 s

Z-order by or order by all columns:

Table row count optimize time
conn_order 10,000,000,000 1515.298 s
conn_zorder 10,000,000,000 11057.194 s

3.1.2. Z-order by benchmark result

By querying the tables before and after optimization, we find that:

10 billion data and 200 files and Query resource: 200 core 600G memory

Table Average File Size Scan row count Average query time row count Skipping ratio
conn_random_parquet 1.2 G 10,000,000,000 27.554 s 0.0%
conn_zorder_only_ip 890 M 43,170,600 2.459 s 99.568%
conn_zorder 890 M 54,841,302 3.185 s 99.451%

10 billion data and 1000 files and Query resource: 200 core 600G memory

Table Average File Size Scan row count Average query time row count Skipping ratio
conn_random_parquet 234.8 M 10,000,000,000 27.031 s 0.0%
conn_zorder_only_ip 173.9 M 53,499,068 3.120 s 99.465%
conn_zorder 174.0 M 35,910,500 3.103 s 99.640%

1 billion data and 10000 files and Query resource: 10 core 40G memory

Table Average File Size Scan row count Average query time row count Skipping ratio
conn_random_parquet 2.7 M 1,000,000,000 76.772 s 0.0%
conn_zorder_only_ip 2.1 M 406,572 3.963 s 99.959%
conn_zorder 2.2 M 387,942 3.621s 99.961%