Hortonworks.Cluster Config Guide

Community Driven Apache Hadoop

Apache Hadoop
Cluster Configuration Guide
April 2013

© 2013 Hortonworks Inc.
http://www.hortonworks.com

Apache Hadoop Cluster Configuration Guide © Hortonworks Inc. 2013 Page 2
Introduction
Sizing a Hadoop cluster is important, as the right resources will allow you to optimize the
environment for your purpose. However, this is no simple task as optimizing a distributed
environment and its related software can have its complexities.
We have packaged up years of experience building and optimizing Hadoop cluster into a single
document and present to you this guide that should help guide you through this process. It is
important to note that while we outline suggested sizing configurations, no two implementations
are alike and mileage may vary. To this end, this guide provides you a basic framework to
address sizing. We hope it will be helpful to you.
The Basics
Hadoop and HBase clusters have two types of machines:
• masters (the HDFS NameNode, the MapReduce JobTracker, and the HBase Master)
• and slaves (the HDFS DataNodes, the MapReduce TaskTrackers, and the HBase
RegionServers).
The DataNodes, TaskTrackers, and
HBase RegionServers are co-located or
co-deployed for optimal data locality. In
addition, HBase requires the use of a
separate component (ZooKeeper) to
manage the HBase cluster.
Hortonworks recommends separating
master and slave nodes because of the
following reasons:
• Task workloads on the slave
nodes should be isolated from
the masters.
• Slaves nodes are frequently
decommissioned for
maintenance.

Clusters of three or more machines typically use a dedicated NameNode/JobTracker and all the
other nodes act as the slave nodes. For a medium (45-150 node) to large(300+) cluster, we
recommend expanding this to 4 master nodes as follows:
1. Name Node
2. JobTracker Node
3. Secondary Name Node, HBase Master, Hive Server
* Note – in a half-rack cluster, this would be combined with the Job Tracker
4. Management Node (Ambari, Ganglia, Nagios)
* Note – in a half-rack cluster, this would be combined with the Name Node
For NameNode failover, we recommend setting
up an NFS share to hold the block map that is
accessible through the same mount point on all
of the master nodes above.
That leaves the remaining hosts for Slave
nodes, each running a DataNode, TaskTracker,
and HBase Region Server. As an option, clients
may be installed on these nodes to facilitate
data and processing integration between the
Hadoop cluster and legacy systems within the
environment. This can be seen in the Assign
Slaves and Clients step of the Ambari Cluster
Install Wizard.
Balanced Value and Optimum Price/Performance
A Hadoop cluster is different from common enterprise IT server infrastructure. First, a workload
or job within a Hadoop cluster will have little or no idle time during its life. Second, Hadoop
requires a “balanced value” which guides you to select hardware configurations for master and
slave nodes where consumption of CPU, memory and disk resources all peak together. The
result of a balanced value design is optimum price/performance of your Hadoop cluster.

Hadoop workloads are different from common IT applications
The Hadoop workload is consumes data and creates new data sets and is typically limited by
CPU, memory and disk bandwidth. If you compare this to a typical IT workload you find the
Hadoop workload spends very little time in an idle state. Hence, A Hadoop workloads is typically
a single execution task. Running multiple Hadoop workloads (in version 1.0) on a common set of
hardware often results in saturation of the CPU, memory and/or disk I/O resources. The net is
overall job load loses efficiency due to bottlenecks. We recommend running master and slave
node on physical hardware for the best efficiency and price/performance optimization.
A “shared nothing” architecture promotes scalability
Another key design concept in Hadoop cluster design is “shared nothing”. This concept says
that a master or slave node should not share any hardware resource with any other node. This
design concept highlights the performance issues that will result of using SANs for disk storage
and unified I/O fabrics found in blade servers. Share resource infrastructure assumes that the
average of the demand from the jobs will not exceed the total bandwidth of the resource. The
performance issues come from the non-interactive job profile of Hadoop workloads. Since all
Hadoop jobs are demanding max performance the shared resources will become saturated and
result in a performance bottleneck.
The benefit of a “shared nothing” architecture comes when nodes are added to a cluster. Since
each node is not sharing any resource with other nodes the addition of a new node results in
added CPU, memory and disk I/O resources. When we couple this with the ability of a Hadoop
cluster to break a larger job into multiple independent jobs the results is a near linear scalability
of the cluster. If you need to double the capacity of the cluster you can easily double the number
of nodes.
Balance: the key to high performance & affordable Hadoop Clusters
When building Hadoop Clusters make sure all components are balanced in capacity, cost and
performance. With a balanced cluster we minimize performance bottlenecks and avoid excess
unused capacity. Since there are many more slave nodes than master nodes most of our
discussion will revolve around slave node design consideration.

Common Cluster Component Guidelines
If we look at a node design for master and slave nodes we see we have control over selection of
CPU, memory and disk I/O. We have secondary considerations of power, cooling and physical
space. We need to keep these in mind as the size of the cluster scales up it acts as a multiplier
and these secondary considerations can very quickly make our design impractical to implement.
• CPU and memory selection
Always start with CPU selection. Here we need to look at price/performance of our CPU
choices. Picking the current top of the line CPUs will result in a very expensive cluster as
you multiple out this cost by the number of nodes. Going the other direction picking
underpowered CPUs results in you cluster not being able to meet its performance goals.
Since there is strong innovation happening in the CPU market the sweet spot will move
quarter to quarter. Right now we find the Intel Sandybridge quad core processors to be
good value. We expect that the hex core processors will become more competitive and
when you do your design you should check on the price/performance ratio.
• Disk selection
The amount of disk on each slave node will define the amount of data your cluster can
hold. The speed of the disks will be one of the bottlenecks that will constrain the slave
node performance. So our disk selection is going to be a tradeoff between performance
and capacity. The highest capacity disks us a SATA interface. SAS drives have a more
sophisticated command set which will give you better performance and reliability but
SAS drives cost more. We have found that near line SAS drives gives us a good
combination of both capacity and performance. A near line SAS drive is the marriage of
a SAS interface to a SATA disk mechanism. So we get the capacity of SATA with some
of the performance of SAS. We do lose on reliability and life span but since Hadoop is
designed to accept drive failures this is an acceptable trade off. This kind of design
decision is typical of what you will do as you iterate through your cluster design.
• Networking
Networking for Hadoop clusters is simple since most of the data traffic is taking place
inside the slave nodes. The network is used for control, bringing in new data and moving
out the results of jobs. For each rack of servers there will be a 1G rack level switch with
a 10G uplink. that connects to the nodes in the rack. For small to medium clusters we

can just daisy chain these switches together. When you go over 5-7 racks you will then
need a 10GB spine switches to interconnect the racks.
Our trade off decision in this case is 1 GB Ethernet vs 10 GB Ethernet. This decision will
be made on cost and infrastructure considerations. While the incremental cost of 10 GB
Ethernet interfaces may be reasonable the next question is can the interconnect
infrastructure handle it?
Later we will talk about cluster isolation and if we have isolated the cluster this may be a
good decision. From a balanced value point of view right now 1 GB Ethernet between
nodes in a rack with a 10GB Ethernet backbone will meet most current designs.
• Redundancy in cluster components
In the design of an Apache Hadoop cluster we apply redundancy sparingly. From a
cluster point of view the built in redundancy of Hadoop software allows a cluster to lose
a drive, a slave node or if the scale is large enough a whole rack of servers. So as your
cluster grows the loss of any component diminishes. This is part of the beauty of the
Hadoop design.
The important place to apply hardware redundancy, i.e. extra power supplies, multiple
network paths, etc. is with the master nodes. This goes back to the fact that the loss of
a master node will affect the whole cluster. This is the place to employ enterprise grade
server hardware to minimize the chance of losing a master node.
The exact opposite happens with the slave nodes. Since data is replicated through the
cluster the loss of a disk drive or a slave node it managed by Apache Hadoop. In a large
enough cluster replacement of failed hardware is a scheduled maintenance task and
Apache Hadoop manages replacement of hardware so there is no cluster downtime. It is
important to realize that slave hardware failure is not a priority one issue and it can be
resolved the next time there is a hardware replacement maintenance event.
Network redundancy changes as you go from small to large clusters. In a small and
medium cluster loss of a switch has a larger impact than in a large cluster. We do not
recommend redundant switches but we do recommend redundant paths so you can
minimize the impact of a switch failure. We mention this later on but a good design

pattern to use here is to physically split the cluster into two equal halves. So if one
cluster has a failure the other cluster can still process jobs.
Master Node Design Guidelines
The master nodes are key components of your Hadoop cluster. Since there are only a few of
them we can relax some of our cost constraints. We want to optimize these nodes for
availability. We recommend using enterprise quality hardware since a master node failure will
impact the cluster. Redundant power supplies and network connections are applicable here to
help increase the availability of the node.
• Memory
Memory demand for a master node is based on the NameNode data structures that
grow with the storage capacity of your cluster. We found 1 GB per petabyte of storage is
a good guideline for master node memory. You then need to add on your OS overhead,
etc. We have found that with Intel Sandybridge processors 32GB is more than enough
memory for a master node.
• Local disk configuration
The local disks on a master node should be configured using RAID10 with hot spares.
This configuration is picked to give good performance and fast recovery on loss of a
disk. We recommend 6 x 1TB drives. Four are allocated to the Raid 10 with two hot
spares.
• NFS server for cluster meta-data
The critical element of the master nodes are the meta-data files that hold the information
about the cluster. Loss of these files will render your cluster useless. A best practice is
the place all these files on a small NFS server. This will allow you to quickly recover from
a loss of a master node by simply starting a new copy of the master node and pointing it
at the metadata files on the NFS share.
• HA considerations
In order to protect the master nodes many clusters use an HA configuration to maintain
the availability of the master nodes. However if we look at the rate of failure of master
nodes, especially if they are built with redundant hardware this does not happen very

often. So you need to balance the complexity of HA against the possibility of failure of
the master node. Also if you implement HA it needs to be tested regularly to make sure it
works and you know how to transition through a failover. Overall you need to consider
the trade off of complexity vs. the recovery benefits from HA.
An alternate configuration would be to split your servers into two independent clusters
and replicate the data between the two clusters. In the beginning you keep the two
clusters together until operations has stabilized. Now when a cluster fails you can just
move jobs over to the other cluster. You are also set up very nicely to later move one
cluster to a remote datacenter to achieve geographical dispersion. The penalty to this
configuration is the data replication and you cannot use all your slave nodes on one big
problem.
• NameNode guideline
An important resource for NameNodes is memory to hold working data structures. A
very rough guideline for NameNode memory is there needs to be 1GB of RAM for every
petabyte of allocated storage space.
Slave node design guidelines
The design of slave nodes is dominated by the fact there are very many of them. Anything
you do good or bad is multiplied by the number of nodes. As noted above you are looking to
balance the CPU, memory and disk I/O configuration such that they all reach their peak
limits around the same point.
• Memory
If we follow the Intel Sandybridge processor selection we have found 32 GB of ram to be
sufficient for a dual quad core server. Remember that memory is cheap but when you
multiple the cost of the number of slave nodes the extra money spent on ram could
allow you to buy a few more slave nodes.
• Disk
Disk selection is limited by the I/O bandwidth of the processor chip set and the disk
controller card. Based on our Intel Sandybridge recommendation we find that 6 x 1TB
disks configured as JBOD will do well. A JBOD configuration keeps the disk controller
out of the way and allow for full speed I/O with the processor.

Cluster Design Tradeoffs
We classify clusters as small (around 2-3 racks), medium(4-10 racks) and large(above 10 racks).
What we have been covering so far are design guidelines and part of the design process is to
understand how to bend the design guidelines to meet you goals. In the case of small, medium
and large clusters things get progressively more stringent and sensitive when you bend the
guidelines. For a small the smaller number of slave nodes allow you greater flexibility in your
decisions. There are a few guidelines you don’t want to violate like isolation.
When you get to a medium size cluster the number of nodes will increase your design sensitivity.
You also now have enough hardware the physical plant issues of cooling and power become
more important. Your interconnects also become more important. At the large scale things
become really sensitive and you have to be careful because making a mistake here could result
in a design that will fail. Our experience at Hortonworks has allowed us to develop expertise in
this area and we strongly recommend you work with us if you want to build Internet scale
clusters.
Growing Your Cluster
Once you get your initial cluster built your success will create a demand to increase your cluster
size. This will be adding new slave nodes. We recommend adding to your cluster in increments
of racks of nodes. This keeps things simpler since you just need to replicate racks. You will
probably be expanding in your existing datacenter so try to keep the new racks near because
you will need to directly connect the new racks to the cluster.
As you grow from a small to medium cluster you want to watch the bandwidth of the rack
switches. At this point the racks are daisy chained together. When you exceed 5-7 racks you
may need to introduce a spine switch that interconnects the rack switches. Growing the cluster
one rack at a time is a good strategy because it will let you spot issues as they emerge vs have a
large change in the cluster and trying to find issues. Otherwise as you grow to the 10-14 rack
range the cluster should grow nicely.
When you go from medium to large clusters your physical infrastructure, space, power, cooling,
etc. will be the driving functions. As server best practices dictate test as you build. Integrate and
test each rack or row. Make sure you are adding a solid and stable resource to your cluster.

If you are planning a major expansion please engage your Hortonworks team as we have been
through many cluster expansions. We have the knowledge and tools to make this a smooth
transition.
Isolation
A Hadoop cluster is an integrated supercomputer cluster. As such issues like network
interconnect latency is important. It is strongly recommended that your Hadoop cluster be
implemented on an isolated network with a minimal amount of network complexity. We have
installations where the Hadoop cluster servers were scattered through the datacenter and the
result was network instability of the whole datacenter. You need to isolate and localize the
Hadoop servers from the rest of the datacenter.
Heterogenous Slave Nodes
In your first build out of your cluster it will be homogenous and all the slave nodes will be the
same or quite similar. Over time technology changes and you will by buying different node
configurations.
When you have a mix of faster and slower nodes Apache Hadoop does make adjustments in the
job stream. A feature of Apache Hadoop is called speculative execution. This happens when a
node is behind in processing its share of a job. The job tracker will start another copy of that
task on another node just in case the first one fails. It will then take the results of whichever node
finishes first. So this works our fine at the job level.
At the cluster level this results in some inefficiency because you are running a task twice. Where
problems might results is when there a large performance differences in nodes. Some things to
watch for are things like SATA I/SATA II disks, 1GB/10GB Ethernet, difference in spindle count,
ext3/xfs file systems, etc. These things make a difference in node performance and what can
happen is that your cluster will develop a hot spot where a server or set of servers become a
bottleneck in getting jobs done. In general you want to keep slave nodes balanced so that any
node can run a task equally well.
Small Hadoop Cluster Design Example

A small Hadoop cluster could consist of a two-level architecture built with rack-mounted servers
housed in one or two racks. With a high density of disks (6-12 per node) on the slave nodes,
Hortonworks does recommend two racks to start in order to best support the load, power, &
cooling specifications of most datacenters as well as the expansion that most customers
undertake after their initial Hadoop implementations – so these constraints / plans should be
reviewed before choosing a single rack. A single rack of servers should be interconnected using
a 1 Gigabit Ethernet (GbE) switch. For purposes of this cluster planning document, we are
assuming a single rack is 42U.
Recommended hardware specifications for a half-rack cluster are as follows:
Rack 1: Small Cluster
Master Nodes
2 Servers
2 - Intel Quad core Sandybridge processors
8 - 8GB DDR3 rDIMMs (Sparing, ECC)
1 – Perc S100 (RHEL6/xfs – RAID10)
6 – 1 TB, NL-SAS disks
Built-in dual GbE
Dual Power Supply
Slave Node
8 x 2U servers
Intel Quad core Sandybridge processor
8 - 4GB DDR3 DIMMs
2 – Internal 2.5” SATA drives for root (RHEL6/xfs)
6 – 2TB, SATA disks (SATA/ahci)
Built-in dual GbE
Dual Power Supply
Rack-level Switch: 1U 24 – ports GbE
This half-rack cluster is sized for 96 TB of Raw Storage. To calculate how much usable disk
space that provides we typically plan for 2 replicas for redundancy purposes with 20%
headroom on each disk. This breaks down to 26.6 TB usable disk across this 8 node cluster.
The calculation is as follows:
Usable Disk = Raw Storage / (3 * 1.2)
Usable Disk = 96 TB / 3.6
Usable Disk = 26.6 TB

Typical Hadoop Cluster – Multiple Racks
Hortonworks recommends starting with a multiple rack configuration for ease of expansion.
Typically, a large Hadoop cluster consists of a three-level architecture built with rack-mounted
servers. Each rack of servers is interconnected using a 1 Gigabit Ethernet (GbE) switch. Each
rack-level switch is connected to a cluster-level switch (which is typically a larger port-density
10GbE switch). These cluster-level switches may also interconnect with other cluster-level
switches or even uplink to another level of switching infrastructure.
The Master Nodes and Slave Nodes would typically deployed in two racks, inter-connected
through a cluster-level switch (which is typically a larger port-density 10GbE switch). For
purposes of this cluster planning document, we are assuming a single rack is 42U.
Recommended hardware specifications for a large cluster are as follows:

Rack 1: Cluster
Master Nodes
4 x 2U Servers
2 – Intel Sandybridge quad core processors
8 – 4 GB DDR3 rDIMMs (Sparing, ECC)
1 – Perc S100 (RHEL6/xfs – RAID10)
6 – 1 TB, NL-SAS disks
Built-in dual GbE
Dual Power Supply
Slave Node
8 x 2U servers
Intel Quad core Sandybridge processor
8 - 4GB DDR3 DIMMs
Built-in dual GbE
Dual Power Supply
Rack-level Switch: 1U 24 – ports GbE
1 – 10GbE SFP (optional)
Cluster-level Switch: 1U 24 – ports GbE
Rack 2-n: Cluster Data Nodes
Data Nodes
20 x 2U Servers
Intel Sandybridge quad core
12 - 4GB DDR3 DIMMs (2/channel, clocks to 1067, 3 channels/socket)
1 – 12x3.5” Drive Bay Option
Built-in dual GbE
Dual Power Supply
Rack-level Switch: 1U 24 – ports GbE per rack
This multi-rack cluster is sized initially for 240 TB of Raw Storage. To calculate how much usable
disk space that provides we typically plan for 2 replicas for redundancy purposes with 20%
headroom on each disk. This breaks down to 66.7 TB usable disk across a 20-node cluster. The
calculation is as follows:
Usable Disk = Raw Storage / (3 * 1.2)

Usable Disk = 240 TB / 3.6
Usable Disk = 66.7 TB
Additional capacity can be added to this multi-rack cluster by half-rack (10) or full-rack (20)
increments of slave nodes. A half-rack of slave nodes would add 120 TB of additional Raw
Storage (33.4 TB of additional Usable Disk). A full-rack of slave nodes would add 240 TB of
additional Raw Storage (66.7 TB of additional Usable Disk).
Conclusion
The greatest trait of Apache Hadoop cluster design is the larger it gets the more resilient it gets.
When you have a large numbers of nodes the loss of a disk, server or even rack diminishes.
When you make your component selections we have found more nodes are better than more
powerful nodes. Also the less you pay for each node the more nodes you can buy. This directly
translates into how much work you can process with your cluster.
Key Takeway: A recipe for configuring a clean stable cluster
We leave you with a set of guidelines and suggestions on how to configure a clean stable
infrastructure on which to install your Apache Hadoop cluster. These guidelines were developed
by engineers with years of Hadoop experience and present their preferred default configuration.
Following these guidelines will help you avoid 80% of the throughput bottlenecks in a cluster. If
this is your first cluster you should start here, get the cluster stabilized and then start modifying it
for what you are doing. This will get you a known stable baseline system.
• RedHat Enterprise Linux 6.3 – there are improvements in this release for memory page
cache BDI flusher performance Hortonworks Data Platform 1.2 or higher – picks up
fadvise/drop-behind and the 4MB readahead for file system performance
• Use XFS file system on all HDFS LUNs – as of 3/10/13 there is a write bug in ext4, ext4
read performance is fine
• For 2TB disks – create a 250GB temp and a 1.75 HDFS partition, this will isolate HDFS
temp directory churn, on a 12 drive node this gives you 3TB of temp which is sufficient
• Use XFS on the temp directory partitions also

About Hortonworks
Hortonworks is a leading commercial vendor of Apache Hadoop, the preeminent open source platform for storing, managing and analyzing big data. Our distribution,
Hortonworks Data Platform powered by Apache Hadoop, provides an open and stable foundation for enterprises and a growing ecosystem to build and deploy big
data solutions. Hortonworks is the trusted source for information on Hadoop, and together with the Apache community, Hortonworks is making Hadoop more robust
and easier to install, manage and use. Hortonworks provides unmatched technical support, training and certification programs for enterprises, systems integrators
and technology vendors.
3460 West Bayshore Rd.
Palo Alto, CA 94303 USA
US: 1.855.846.7866
International: 1.408.916.4121
www.hortonworks.com
Twitter: twitter.com/hortonworks
Facebook: facebook.com/hortonworks
LinkedIn: linkedin.com/company/hortonworks
• Do not create /hdfs or HDFS temp directory on the root volume
• Slots: For Intel processors a dual processor hex core system will show up as a 24 CPU
system. However half of the “CPUs” are due to the Intel Hyper Threading feature. To
account for this, set the Apache Hadoop map slots to 2/3 of the number of cores.
As noted above these suggestions are all implemented before you install Apache Hadoop. The
next step after this is to implement our Hortonworks Data Platform pre-installation checklist.
Apache Hadoop Cluster design is a serious platform-engineering project and design decisions
need to be well understood. What you can get away with in a small cluster may cause issues as
the cluster grows. This guide has given you the basics of Hadoop Cluster design and you can
easily design a small cluster. We do strongly recommend you engage Hortonworks if you need
to build a large Internet scale cluster. Our proven experience, learned from Yahoo, allows us to
help you avoid common design issues and deliver a stable working Hadoop cluster.

Hortonworks.Cluster Config Guide

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Hortonworks.Cluster Config Guide (20)

More from Douglas Bernardini (20)

Recently uploaded (20)

Hortonworks.Cluster Config Guide