涉及到一些关于硬件的东西，我也不是很懂，记录下来有待以后学习。
Hadoop集群一般都是由小到大，刚开始可能只有4到6个节点，随着存储数据的增加，计算量的增大，内存需求的增加，集群慢慢变大。
比如按照数据存储量增大集群，每个星期数据存储3TB数据，HDFS的block备份数为3，则集群就需要9TB的磁盘，一般还要再预估25%buffer。如果一台机器的存储量为16*3T，则大概每个月往集群中增加1台机器。
如何进行硬件选择？一般Hadoop节点分成管理节点（master node）和工作节点（work node）。管理节点上跑NameNode,Standby NameNode,ResourceManager，SecondaryNameNode。工作节点上跑DataNode,NodeManager,impala server进程。
work nodes的推荐配置：
中级配置（deep storage, 1Gb Ethernet ）：
– 16 x 3TB SATA II hard drives, in a non-RAID, JBOD* configuraGon – 1 or 2 of the 16 drives for the OS, with RAID-1 mirroring
– 2 x 6-core 2.9GHz CPUs, 15MB cache
– 256GB RAM
– 2x1 Gigabit Ethernet
高级配置（high memory, spindle dense, 10Gb Ethernet ）：
– 24 x 1TB Nearline/MDL SAS hard drives, in a non-RAID, JBOD* configuraGon
– 2 x 6-core 2.9GHz CPUs, 15MB cache – 512GB RAM (or more)
– 1x10 Gigabit Ethernet
Work Node不推荐RAID，不推荐Blade Servers。
master node的推荐配置：
Carrier-class hardware
Dual power supplies
Dual Ethernet cards
– Bonded to provide failover
RAIDed hard drives
Reasonable amount of RAM
– 64 GB for clusters of 20 nodes or less
– 96 GB for clusters of up to 300 nodes
– 128 GB for larger clusters
不推荐部署在虚拟化的主机上，因为虚拟化会带了很多不确定性，比如虚拟的三个server,实际的存储可能在一个物理server上，给hdfs的block备份带来风险。
Network推荐：
Nodes are connected to a top-of-rack switch
Nodes should be connected at a minimum speed of 1Gb/sec
Consider 10Gb/sec connecAons in the following cases:
– Clusters storing very large amounts of data
– Clusters in which typical jobs produce large amounts of intermediate
data
Racks are interconnected via core switches
Core switches should connect to top-of-rack switches at 10Gb/sec or faster
Beware of oversubscripAon in top-of-rack and core switches
Consider bonded Ethernet to miAgate against failure
Consider redundant top-of-rack and core switches
用域名，避免用IP地址，最好配DNS.
OS建议选centos or RedHat Enterprise Linux (RHEL)
磁盘划分越多越好，避免LVM(Logical Volume Manager)，设置noatime。
存储的文件size越大越好。
OS,network,system time, user and group和component版本等等的配置，可以通过Cloudera Manager Host Inspector 进行check。