Shulou(Shulou.com)06/03 Report--

Github of Fayson: https://github.com/fayson/cdhproject

It is recommended to follow the official account of Wechat: "Hadoop practice", ID:gh_c4c535955d0f, or scan the QR code at the end of the article.

1. Networking configuration

The recommended Cloudera networking configuration is shown in the following figure, which mainly includes data network and management network.

two。 Data network

A data network is a private cluster data interconnection between nodes used for data access, such as moving data between nodes within a cluster, or importing data into a CDH cluster. CDH clusters are usually connected to data networks within the enterprise.

Two TOR switches are required: one for out-of-band management and one for the data network for CDH. Out-of-band management of nodes requires at least one 1GbE switch. The switch for a data network is generally 10GbE, depending on the workload.

The recommended 1GbE switch is the Lenovo RackSwitch G8052. The 10Gb Ethernet switch can provide additional Imax O bandwidth for better performance. The recommended 10GbE switch is Lenovo System NetworkingRackSwitch 8272.

The two Broadcom 10GbE ports of each node can be bound and connected to the G8272 switch to improve performance or configure HA. The data network can be configured to use VLAN.

Note: Cloudera does not support multi-homing whether it is a worker node or a management node.

3. Hardware management network

The hardware management network is a 1 GbE network for out-of-band hardware management. Through the integrated management module II (IMM2) in the System x3650 M5 server, out-of-band management can achieve hardware-level management of cluster nodes, such as node deployment, basic input / output system (BIOS) configuration, status and power status.

Hadoop does not depend on IMM2. Administrative links can be separated to different VLAN or subnets according to customer requirements. The management network is usually directly connected to the customer's management network.

The reference architecture requires a 1 Gb Ethernet TOR switch for hardware management networks. Administrators can also access all nodes in the cluster through the customer management network, and in the figure in Chapter 1, the management link is connected to a dedicated IMM2 port on the integrated 1 GBaseT adapter.

4. Multi-rack network

The reference architecture configuration of the data network mentioned above is made up of a single network topology. For a multi-rack architecture, a Lenovo RackSwitch G8316 core switch is also required. In this case, the second Broadcom10 GbE port can be connected to the second Lenovo RackSwitch G8272. The overload rate (over-subscription ratio) of G8272 is 1:2.

The following figure shows how to configure a network when a CDH cluster is installed on multiple racks. The G8272 switch in each rack is connected to the core G8316 switch through two aggregated 40 GbE uplinks.

Note: to simplify this diagram, only one G8272 is drawn, but two G8272 are recommended and configured as HA.

40GbE is recommended for cross-rack switches, and Lenovo System NetworkingRackSwitch G8316 can be used. The best practice is to install redundant core switches for each rack to avoid a single point of failure. Within each rack, the G8052 switch can be optionally configured with uplinks with two G8272 switches to allow management VLAN to propagate between cluster racks through the G8316 core switch. For large clusters, Lenovo System NetworkingRackSwitch G8332 is recommended because the price per 40 Gb port is lower than that of G8316. It can be configured so that many racks can access each other's network, but some specific deployment configurations may be required to accommodate fast addressing of more than three racks.

If you start with a multi-rack solution, or if some racks are slowly added as the system expands, we recommend deploying nodes related to CDH management services in separate racks to maximize fault tolerance.

5.CDH other network requirements

Hadoop network requirements:

1. All Hadoop server nodes should be unique networks, and there is no case of sharing the network Imando with nodes of other applications.

two。 Each server should be configured with static IP. If dynamic IP is configured, the IP address of the machine will change when the machine is rebooted or the DNS lease expires, which will lead to Hadoop service failure.

3. Private TOR switch.

4. Dedicated core switching blades or core switches.

5. Try to ensure that the application server is "close" to Hadoop.

6.CDH only supports IPv4, not IPv6

7. The network connection between the racks should be fast enough.

8. Ensure that the network interface should be consistent for all nodes in the cluster. (for example, MTU settings should be the same)

9. Turn off the Huge Page compaction of all nodes

10. Ensure that all network connections in the cluster are monitored, such as collisions and packet loss issues. To facilitate later troubleshooting.

Set your mind for heaven and earth, set your life for the people, continue to learn for the past, and open peace for all eternity.

It is recommended to follow Hadoop practice, the first time, share more Hadoop practical information, welcome to forward and share.

Original article, welcome to reprint, reprint please indicate: reproduced from the official account of Wechat Hadoop

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.