Optimizing Put and Scan Performance

Scenario

HBase has many configuration parameters related to read and write performance. The configuration parameters need to be adjusted based on the read/write request loads. This section describes how to optimize read and write performance by modifying the RegionServer configurations.

This section applies to MRS 3.x and later versions.

Procedure

• JVM GC parameters

  Suggestions on setting the RegionServer GC_OPTS parameter:

  • Set -Xms and -Xmx to the same value based on your needs. Increasing the memory can improve the read and write performance. For details, see the description of hfile.block.cache.size in Table 2 and hbase.regionserver.global.memstore.size in Table 1.
  • Set -XX:NewSize and -XX:MaxNewSize to the same value. You are advised to set the value to 512M in low-load scenarios and 2048M in high-load scenarios.
  • Set X-XX:CMSInitiatingOccupancyFraction to be less than and equal to 90, and it is calculated as follows: 100 x (hfile.block.cache.size + hbase.regionserver.global.memstore.size + 0.05).
  • -XX:MaxDirectMemorySize indicates the non-heap memory used by the JVM. You are advised to set this parameter to 512M in low-load scenarios and 2048M in high-load scenarios.
    

    The -XX:MaxDirectMemorySize parameter is not used by default. If you need to set this parameter, add it to the GC_OPTS parameter.

• Put parameters

  RegionServer processes the data of the put request and writes the data to memstore and HLog.

  • When the size of memstore reaches the value of hbase.hregion.memstore.flush.size, memstore is updated to HDFS to generate HFiles.
  • Compaction is triggered when the number of HFiles in the column cluster of the current region reaches the value of hbase.hstore.compaction.min.
  • If the number of HFiles in the column cluster of the current region reaches the value of hbase.hstore.blockingStoreFiles, the operation of refreshing the memstore and generating HFiles is blocked. As a result, the put request is blocked.

  Table 1 Put parameters

  Parameter	Description	Default Value
  hbase.wal.hsync	Indicates whether each WAL is persistent to disks. For details, see Improving Put Performance.	true
  hbase.hfile.hsync	Indicates whether HFile write operations are persistent to disks. For details, see Improving Put Performance.	true
  hbase.hregion.memstore.flush.size	If the size of MemStore (unit: Byte) exceeds a specified value, MemStore is flushed to the corresponding disk. The value of this parameter is checked by each thread running hbase.server.thread.wakefrequency. It is recommended that you set this parameter to an integer multiple of the HDFS block size. You can increase the value if the memory is sufficient and the put load is heavy.	134217728
  hbase.regionserver.global.memstore.size	Updates the size of all MemStores supported by the RegionServer before locking or forcible flush. It is recommended that you set this parameter to hbase.hregion.memstore.flush.size x Number of regions with active writes/RegionServer GC -Xmx. The default value is 0.4, indicating that 40% of RegionServer GC -Xmx is used.	0.4
  hbase.hstore.flusher.count	Indicates the number of memstore flush threads. You can increase the parameter value in heavy-put-load scenarios.	2
  hbase.regionserver.thread.compaction.small	Indicates the number of small compaction threads. You can increase the parameter value in heavy-put-load scenarios.	10
  hbase.hstore.blockingStoreFiles	If the number of HStoreFile files in a Store exceeds the specified value, the update of the HRegion will be locked until a compression is completed or the value of base.hstore.blockingWaitTime is exceeded. Each time MemStore is flushed, a StoreFile file is written into MemStore. Set this parameter to a larger value in heavy-put-load scenarios.	15

• Scan parameters

  Table 2 Scan parameters

  Parameter	Description	Default Value
  hbase.client.scanner.timeout.period	Client and RegionServer parameters, indicating the lease timeout period of the client executing the scan operation. You are advised to set this parameter to an integer multiple of 60000 ms. You can set this parameter to a larger value when the read load is heavy. The unit is milliseconds.	60000
  hfile.block.cache.size	Indicates the data cache percentage in the RegionServer GC –Xmx. You can increase the parameter value in heavy-read-load scenarios, in order to improve cache hit ratio and performance. It indicates the percentage of the maximum heap (-Xmx setting) allocated to the block cache of HFiles or StoreFiles.	When offheap is disabled, the default value is 0.25. When offheap is enabled, the default value is 0.1.

• Handler parameters

  Table 3 Handler parameters

  Parameter	Description	Default Value
  hbase.regionserver.handler.count	Indicates the number of RPC server instances on RegionServer. The recommended value ranges from 200 to 400.	200
  hbase.regionserver.metahandler.count	Indicates the number of program instances for processing prioritized requests. The recommended value ranges from 200 to 400.	200