Apache
/
hadoop
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

HDFS-3697. Enable fadvise readahead by default. Contributed by Todd Lipcon. 

git-svn-id: https://svn.apache.org/repos/asf/hadoop/common/trunk@1364698 13f79535-47bb-0310-9956-ffa450edef68
This commit is contained in:
Todd Lipcon 2012-07-23 16:27:59 +00:00
parent 9c9f29ac7e
commit b7b8db51de
3 changed files with 77 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt
View File

@ -363,6 +363,8 @@ Branch-2 ( Unreleased changes )
HDFS-3110. Use directRead API to reduce the number of buffer copies in
libhdfs (Henry Robinson via todd)
HDFS-3697. Enable fadvise readahead by default. (todd)
BUG FIXES
HDFS-3385. The last block of INodeFileUnderConstruction is not

2
hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/DFSConfigKeys.java
View File

@ -74,7 +74,7 @@ public class DFSConfigKeys extends CommonConfigurationKeys {
public static final String DFS_DATANODE_BALANCE_BANDWIDTHPERSEC_KEY = "dfs.datanode.balance.bandwidthPerSec";
public static final long DFS_DATANODE_BALANCE_BANDWIDTHPERSEC_DEFAULT = 1024*1024;
public static final String DFS_DATANODE_READAHEAD_BYTES_KEY = "dfs.datanode.readahead.bytes";
public static final long DFS_DATANODE_READAHEAD_BYTES_DEFAULT = 0;
public static final long DFS_DATANODE_READAHEAD_BYTES_DEFAULT = 4 * 1024 * 1024; // 4MB
public static final String DFS_DATANODE_DROP_CACHE_BEHIND_WRITES_KEY = "dfs.datanode.drop.cache.behind.writes";
public static final boolean DFS_DATANODE_DROP_CACHE_BEHIND_WRITES_DEFAULT = false;
public static final String DFS_DATANODE_SYNC_BEHIND_WRITES_KEY = "dfs.datanode.sync.behind.writes";

74
hadoop-hdfs-project/hadoop-hdfs/src/main/resources/hdfs-default.xml
View File

@ -714,6 +714,80 @@
</description>
</property>
<property>
<name>dfs.datanode.readahead.bytes</name>
<value>4193404</value>
<description>
While reading block files, if the Hadoop native libraries are available,
the datanode can use the posix_fadvise system call to explicitly
page data into the operating system buffer cache ahead of the current
reader's position. This can improve performance especially when
disks are highly contended.
This configuration specifies the number of bytes ahead of the current
read position which the datanode will attempt to read ahead. This
feature may be disabled by configuring this property to 0.
If the native libraries are not available, this configuration has no
effect.
</description>
</property>
<property>
<name>dfs.datanode.drop.cache.behind.reads</name>
<value>false</value>
<description>
In some workloads, the data read from HDFS is known to be significantly
large enough that it is unlikely to be useful to cache it in the
operating system buffer cache. In this case, the DataNode may be
configured to automatically purge all data from the buffer cache
after it is delivered to the client. This behavior is automatically
disabled for workloads which read only short sections of a block
(e.g HBase random-IO workloads).
This may improve performance for some workloads by freeing buffer
cache spage usage for more cacheable data.
If the Hadoop native libraries are not available, this configuration
has no effect.
</description>
</property>
<property>
<name>dfs.datanode.drop.cache.behind.writes</name>
<value>false</value>
<description>
In some workloads, the data written to HDFS is known to be significantly
large enough that it is unlikely to be useful to cache it in the
operating system buffer cache. In this case, the DataNode may be
configured to automatically purge all data from the buffer cache
after it is written to disk.
This may improve performance for some workloads by freeing buffer
cache spage usage for more cacheable data.
If the Hadoop native libraries are not available, this configuration
has no effect.
</description>
</property>
<property>
<name>dfs.datanode.sync.behind.writes</name>
<value>false</value>
<description>
If this configuration is enabled, the datanode will instruct the
operating system to enqueue all written data to the disk immediately
after it is written. This differs from the usual OS policy which
may wait for up to 30 seconds before triggering writeback.
This may improve performance for some workloads by smoothing the
IO profile for data written to disk.
If the Hadoop native libraries are not available, this configuration
has no effect.
</description>
</property>
<property>
<name>dfs.client.failover.max.attempts</name>
<value>15</value>