[Hadoop] Hive 性能+特性

[Hadoop] Hive 性能

利用Hive Bulk Inport数据到Hbase http://wiki.apache.org/hadoop/Hive/HBaseBulkLoad

生成测试数据

/home/bmb/jdk1.6.0_16/bin/java  -cp examples.zip examples.CreateLogFile 1 1000000

/home/bmb/jdk1.6.0_16/bin/java  -cp examples.zip examples.CreateLogFile 1000000 2000000

/home/bmb/jdk1.6.0_16/bin/java  -cp examples.zip examples.CreateLogFile 2000000 3000000

创建性能测试表

不带压缩的测试表

drop table  p_test_data;

CREATE TABLE p_test_data (
  id INT,
  content STRING,
  time  STRING)
ROW FORMAT DELIMITED
FIELDS TERMINATED BY ','
STORED AS TEXTFILE;

导入数据

LOAD DATA LOCAL INPATH '/home/iic/hadoop-0.20.2/test_0_1000000.log' INTO TABLE p_test_data;

LOAD DATA LOCAL INPATH '/home/iic/hadoop-0.20.2/test_1000000_2000000.log' INTO TABLE p_test_data;

LOAD DATA LOCAL INPATH '/home/iic/hadoop-0.20.2/test_2000000_3000000.log' INTO TABLE p_test_data;

set mapred.reduce.tasks=1;

select count(a.id) from p_test_data a;

Time taken: 27.265 seconds

select a.id,a.content,a.time from p_test_data a where a.id=1;

Time taken: 18.086 seconds

INSERT OVERWRITE DIRECTORY '/tmp/p_test_data_out'

select a.time,count(1) from p_test_data a group by a.time;

Time taken: 32.899 seconds

带压缩的测试表

(框架检测到输入文件的后缀是.gz和.lzo，就会使用对应的CompressionCodec自动解压缩这些文件 )

drop table  p_com_test_data;

CREATE TABLE p_com_test_data (
  id INT,
  content STRING,
  time  STRING)
ROW FORMAT DELIMITED
FIELDS TERMINATED BY ','
STORED AS TEXTFILE;

tar cvf 0_100W.tar test_0_1000000.log

gzip 0_100W.tar

tar cvf 100_200W.tar test_1000000_2000000.log

gzip 100_200W.tar

tar cvf 200_300W.tar test_2000000_3000000.log

gzip 200_300W.tar

导入数据

LOAD DATA LOCAL INPATH '/home/iic/hadoop-0.20.2/0_100W.tar.gz' INTO TABLE p_com_test_data;

LOAD DATA LOCAL INPATH '/home/iic/hadoop-0.20.2/100_200W.tar.gz' INTO TABLE p_com_test_data;

LOAD DATA LOCAL INPATH '/home/iic/hadoop-0.20.2/200_300W.tar.gz' INTO TABLE p_com_test_data;

select a.time,count(1) from p_com_test_data a group by a.time;

Time taken: 26.31 seconds

 此例子是针对小量文件的压缩和不压缩的性能测试，虽然不代表最终结果，但是从本次测试可以发现，压缩的效率更高，

可能是因为压缩文件是作为整个Block给Map，减少了InputSplit的检测和分析。