Configuring Racks for Hosts

Scenario

All hosts in a large cluster are usually deployed on multiple racks. Hosts on different racks communicate with each other through switches. The network bandwidth between different hosts on the same rack is much greater than that on different racks. In this case, plan the network topology based on the following requirements:

• To improve the communication speed, it is recommended that data be exchanged between hosts on the same rack.
• To improve the fault tolerance capability, distribute processes or data of distributed services on different hosts of multiple racks as dispersedly as possible.

Hadoop uses a file directory structure to represent hosts.

The HDFS cannot automatically determine the network topology of each DataNode in the cluster. You need to set the rack name to identify the rack where the host is located so that the NameNode can draw the network topology of the required DataNodes and back up data of the DataNodes to different racks. Similarly, Yarn needs to obtain rack information and allocate tasks to different NodeManagers as required.

If the cluster network topology changes, you need to reallocate racks for hosts on FusionInsight Manager so that related services can be automatically adjusted.

Impact on the System

If the name of the host rack is changed, storage policy for HDFS replicas, Yarn task assignment, and storage location of Kafka partitions will be affected. After the modification, you need to restart the HDFS, Yarn, and Kafka for the configuration to take effect.

Improper rack configuration will unbalance loads (including CPU, memory, disk, and network) among nodes in the cluster, which decreases the cluster reliability and stability. Therefore, before allocating racks, take all aspects into consideration and properly set racks.

Rack Allocation Policies

Physical rack: indicates the real rack where the host resides.

Logical rack: indicates the rack name of the host on FusionInsight Manager.

Policy 1: Each logical rack has nearly the same number of hosts.

Policy 2: The name of the logical rack of the host must comply with that of the physical rack to which the host belongs.

Policy 3: If there are only few hosts on a physical rack, combine this physical rack and other physical racks with few hosts into a logical rack, which complies with policy 1. Hosts in two equipment rooms cannot be placed in one logical rack. Otherwise, performance problems may be caused.

Policy 4: If there are lots of hosts on a physical rack, divide these hosts into multiple logical racks, which complies with policy 1. Hosts with great differences should not be placed in the same logical rack. Otherwise, the cluster reliability will be decreased.

Policy 5: You are advised to set default or other values for logical racks on the first layer, and the values in the same cluster must be consistent.

Policy 6: The number of hosts in each rack cannot be less than 3.

Policy 7: A cluster can contain at most 50 logical racks. If there are too many logical racks in a cluster, the maintenance is difficult.

Best Practices

For example, in a cluster, 100 hosts are located in two equipment rooms A and B. A has 40 hosts and B has 60 hosts. In room A, there are 11 hosts on physical rack Ra1 and 29 hosts on physical rack Ra2. In room B, there are six hosts on physical rack Rb1, 33 hosts on physical rack Rb2, 18 hosts on physical rack Rb3, and three hosts on physical rack Rb4.

According to the rack allocation policy, each logical rack contains nearly the same number (for example, 20) of hosts. The allocation details are as follows:

• Logical rack /default/racka1: 11 hosts on physical rack Ra1 and nine hosts on physical rack Ra2
• Logical rack /default/racka2: the remaining 20 hosts (except the nine hosts of logical rack /default/racka1) on physical rack Ra2
• Logical rack /default/rackb1: six hosts on physical rack Rb1 and 13 hosts on physical rack Rb2
• Logical rack /default/rackb2: the remaining 20 hosts on physical rack Rb2
• Logical rack /default/rackb3: 18 hosts on physical rack Rb3 and three hosts on physical rack Rb4

Rack allocation example:

Procedure

1. Log in to FusionInsight Manager.
2. Click Hosts.
3. Select the check box of the target host.
4. Select Set Rack from the More drop-down list.
   • Set rack names in hierarchy based on the actual network topology. Separate racks from different layers using slashes (/).
   • Rack naming rules are as follows: /level1/level2/... The number of levels must be at least 1, and the name cannot be empty. A rack can contain letters, digits, and underscores (_) and cannot exceed 200 characters.

     For example, /default/rack0.

   • If the hosts in the rack to be modified contain DataNode instances, ensure that the rack name levels of the hosts where all DataNode instances reside are the same. Otherwise, the configuration fails to be delivered.

5. Click OK.