Rack Awareness

Hadoop components are rack-aware. For example, HDFS block placement will use rack awareness for fault tolerance by placing one block replica on a different rack. This provides data availability in the event of a network switch failure or partition within the cluster.

Hadoop master daemons obtain the rack id of the cluster slaves by invoking either an external script or java class as specified by configuration files. Using either the java class or external script for topology, output must adhere to the java org.apache.hadoop.net.DNSToSwitchMapping interface. The interface expects a one-to-one correspondence to be maintained and the topology information in the format of ‘/myrack/myhost’, where ‘/’ is the topology delimiter, ‘myrack’ is the rack identifier, and ‘myhost’ is the individual host. Assuming a single /24 subnet per rack, one could use the format of ‘/192.168.100.0/192.168.100.5’ as a unique rack-host topology mapping.

To use the java class for topology mapping, the class name is specified by the topology.node.switch.mapping.impl parameter in the configuration file. An example, NetworkTopology.java, is included with the hadoop distribution and can be customized by the Hadoop administrator. Using a Java class instead of an external script has a performance benefit in that Hadoop doesn’t need to fork an external process when a new slave node registers itself.

If implementing an external script, it will be specified with the topology.script.file.name parameter in the configuration files. Unlike the java class, the external topology script is not included with the Hadoop distribution and is provided by the administrator. Hadoop will send multiple IP addresses to ARGV when forking the topology script. The number of IP addresses sent to the topology script is controlled with net.topology.script.number.args and defaults to 100. If net.topology.script.number.args was changed to 1, a topology script would get forked for each IP submitted by DataNodes and/or NodeManagers.

If topology.script.file.name or topology.node.switch.mapping.impl is not set, the rack id ‘/default-rack’ is returned for any passed IP address. While this behavior appears desirable, it can cause issues with HDFS block replication as default behavior is to write one replicated block off rack and is unable to do so as there is only a single rack named ‘/default-rack’.

An additional configuration setting is mapreduce.jobtracker.taskcache.levels which determines the number of levels (in the network topology) of caches MapReduce will use. So, for example, if it is the default value of 2, two levels of caches will be constructed – one for hosts (host -> task mapping) and another for racks (rack -> task mapping). Giving us our one-to-one mapping of ‘/myrack/myhost’.