Planning HDFS Capacity

Updated on 2024-10-25 GMT+08:00

View PDF

In HDFS, DataNode stores user files and directories as blocks, and file objects are generated on the NameNode to map each file, directory, and block on the DataNode.

The file objects on the NameNode require certain memory capacity. The memory consumption linearly increases as more file objects generated. The number of file objects on the NameNode increases and the objects consume more memory when the files and directories stored on the DataNode increase. In this case, the existing hardware may not meet the service requirement and the cluster is difficult to be scaled out.

Capacity planning of the HDFS that stores a large number of files is to plan the capacity specifications of the NameNode and DataNode and to set parameters according to the capacity plans.

Capacity Specifications

NameNode capacity specifications

Each file object on the NameNode corresponds to a file, directory, or block on the DataNode.

A file uses at least one block. The default size of a block is 134,217,728, that is, 128 MB, which can be set in the dfs.blocksize parameter. By default, a file whose size is less than 128 MB occupies only one block. If the file size is greater than 128 MB, the number of occupied blocks is the file size divided by 128 MB (Number of occupied blocks = File size/128). The directories do not occupy any blocks.

Based on dfs.blocksize, the number of file objects on the NameNode is calculated as follows:

**Table 1** Number of NameNode file objects
Size of a File	Number of File Objects
< 128 MB	1 (File) + 1 (Block) = 2
> 128 MB (for example, 128 GB)	1 (File) + 1,024 (128 GB/128 MB = 1,024 blocks) = 1,025

The maximum number of file objects supported by the active and standby NameNodes is 300,000,000 (equivalent to 150,000,000 small files). dfs.namenode.max.objects specifies the number of file objects that can be generated in the system. The default value is 0, which indicates that the number of generated file objects is not limited.

DataNode capacity specifications

In HDFS, blocks are stored on the DataNode as copies. The default number of copies is 3, which can be set in the dfs.replication parameter.

The number of blocks stored on all DataNode role instances in the cluster can be calculated based on the following formula: Number of HDFS blocks x 3 Average number of saved blocks = Number of HDFS blocks x 3/Number of DataNodes

**Table 2** DataNode specifications
Item	Specifications
Maximum number of blocks supported by a DataNode instance	5,000,000
Maximum number of blocks supported by a disk on a DataNode instance	500,000
Minimum number of disks required when the number of blocks supported by a DataNode instance reaches the maximum	10

**Table 3** Number of DataNodes
Number of HDFS Blocks	Minimum Number of DataNode Roles
10,000,000	10,000,000 *3/5,000,000 = 6
50,000,000	50,000,000 *3/5,000,000 = 30
100,000,000	100,000,000 *3/5,000,000 = 60

Setting Memory Parameters

Configuration rules of the NameNode JVM parameter

Default value of the NameNode JVM parameter GC_OPTS:

-Xms2G -Xmx4G -XX:NewSize=128M -XX:MaxNewSize=256M -XX:MetaspaceSize=128M -XX:MaxMetaspaceSize=128M -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:CMSInitiatingOccupancyFraction=65 -XX:+PrintGCDetails -Dsun.rmi.dgc.client.gcInterval=0x7FFFFFFFFFFFFFE -Dsun.rmi.dgc.server.gcInterval=0x7FFFFFFFFFFFFFE -XX:-OmitStackTraceInFastThrow -XX:+PrintGCDateStamps -XX:+UseGCLogFileRotation -XX:NumberOfGCLogFiles=10 -XX:GCLogFileSize=1M -Djdk.tls.ephemeralDHKeySize=3072 -Djdk.tls.rejectClientInitiatedRenegotiation=true -Djava.io.tmpdir=${Bigdata_tmp_dir}

The number of NameNode files is proportional to the used memory size of the NameNode. When file objects change, you need to change -Xms2G -Xmx4G -XX:NewSize=128M --XX:MaxNewSize=256M in the default value. The following table lists the reference values.

**Table 4** NameNode JVM configuration
Number of File Objects	Reference Value
10,000,000	-Xms6G -Xmx6G -XX:NewSize=512M -XX:MaxNewSize=512M
20,000,000	-Xms12G -Xmx12G -XX:NewSize=1G -XX:MaxNewSize=1G
50,000,000	-Xms32G -Xmx32G -XX:NewSize=3G -XX:MaxNewSize=3G
100,000,000	-Xms64G -Xmx64G -XX:NewSize=6G -XX:MaxNewSize=6G
200,000,000	-Xms96G -Xmx96G -XX:NewSize=9G -XX:MaxNewSize=9G
300,000,000	-Xms164G -Xmx164G -XX:NewSize=12G -XX:MaxNewSize=12G

Configuration rules of the DataNode JVM parameter

Default value of the DataNode JVM parameter GC_OPTS:

The average number of blocks stored in each DataNode instance in the cluster is: Number of HDFS blocks x 3/Number of DataNodes. If the average number of blocks changes, you need to change -Xms2G -Xmx4G -XX:NewSize=128M -XX:MaxNewSize=256M in the default value. The following table lists the reference values.

**Table 5** DataNode JVM configuration
Average Number of Blocks in a DataNode Instance	Reference Value
2,000,000	-Xms6G -Xmx6G -XX:NewSize=512M -XX:MaxNewSize=512M
5,000,000	-Xms12G -Xmx12G -XX:NewSize=1G -XX:MaxNewSize=1G

Xmx specifies memory which corresponds to the threshold of the number of DataNode blocks, and each GB memory supports a maximum of 500,000 DataNode blocks. Set the memory as required.

Viewing the HDFS Capacity Status

NameNode information
Log in to FusionInsight Manager, choose Cluster > Services > HDFS > NameNode(Active), and click Overview to view information like the number of file objects, files, directories, and blocks in HDFS in Summary area.
DataNode information
Log in to FusionInsight Manager, choose Cluster > Services > HDFS > NameNode(Active), and click DataNodes to view the number of blocks on all DataNodes that report alarms.
Alarm information
Check whether the alarms whose IDs are 14007, 14008, and 14009 are generated and change the alarm thresholds as required.