The first question people raise when getting started with
Hadoop is about selecting appropriate hardware for their Hadoop cluster.
This blog post describes the various factors that Hadoop
administrators take into account. We encourage others to chime in with
their experience configuring production Hadoop clusters. Although
Hadoop is designed to run on industry standard hardware, recommending an
ideal cluster configuration is not as easy as just delivering a list
of hardware specifications. Selecting hardware that provides the best
balance of performance and economy for a given workload requires testing
and validation. For example, users with IO-intensive workloads will
invest in more spindles per core. In this blog post we’ll discuss
workload evaluation and the critical role it plays in hardware
selection.
Marrying storage and compute
Over the past decade IT organizations have standardized on blades and
SANs (Storage Area Networks) to satisfy their grid and
processing-intensive workloads. While this model makes a lot of sense
for a number of standard applications such as web servers, app servers,
smaller structured databases and simple ETL (Extract, Transform, Load)
the requirements for infrastructure has been changing as the amount of
data and number of users has grown. Web servers now have caching tiers,
databases have gone massively parallel with local disk, and ETL jobs are
pushing more data than they can handle locally. Hardware vendors have
created innovative systems to address these requirements including
storage blades, SAS (Serial Attached SCSI) switches, external SATA
arrays and larger capacity rack units.
Hadoop was designed based on a new approach to storing and processing
complex data. Instead of relying on a SAN for massive storage and
reliability then moving it to a collection of blades for processing,
Hadoop handles large data volumes and reliability in the software tier.
Hadoop distributes data across a cluster of balanced machines and uses
replication to ensure data reliability and fault tolerance. Because
data is distributed on machines with compute power, processing can be
sent directly to the machines storing the data. Since each machine in a
Hadoop cluster both stores and processes data, they need to be
configured to satisfy both data storage and processing requirements.
Why workloads matter
In nearly all cases, a MapReduce job will either encounter a
bottleneck reading data from disk or from the network (known as an
IO-bound job) or in processing data (CPU-bound). An example of an
IO-bound job is sorting, which requires very little processing (simple
comparisons) and a lot of reading and writing to disk. An example of a
CPU-bound job is classification, where some input data is processed in
very complex ways to determine an ontology.
Here are several more examples of IO-bound workloads:
- Indexing
- Searching
- Grouping
- Decoding/decompressing
- Data importing and exporting
Here are several more examples of CPU-bound workloads:
- Machine learning
- Complex text mining
- Natural language processing
- Feature extraction
Since our customers need to understand their workloads in order to
fully optimize their Hadoop hardware, we often start out with a classic
chicken and egg problem on our hands. Most teams looking to build a
Hadoop cluster don’t yet know the profile of their workload and often
the first jobs that an organization runs with Hadoop are far different
than the jobs that Hadoop is used for as they build proficiency.
Additionally, some workloads might be bound in unforeseen ways. For
example, sometimes theoretical IO-bound workloads might actually be
CPU-bound because of a user’s choice of compression. Or sometimes
different implementations of an algorithm might change how the MapReduce
job is constrained. For these reasons it makes sense to invest in a
balanced Hadoop cluster when the team is unfamiliar with the types of
jobs they are going to run. The team can benchmark MapReduce jobs once
they’re running on the balanced cluster, to understand how they’re
bound.
It is straightforward to measure live workloads and determine
bottlenecks by putting thorough monitoring in place on the Hadoop
cluster. We recommend installing Ganglia on all Hadoop machines to
provide real-time statistics about CPU, disk, and network load. With
Ganglia installed a Hadoop administrator can then run their MapReduce
jobs and check the Ganglia dashboard to see how each machine is
performing.
In addition to building out a cluster appropriate for the workload,
we encourage our customers to work with a hardware vendor and understand
the economics of power and cooling. Since Hadoop runs on tens,
hundreds, or thousands of nodes an operations team can save a
significant amount of money by investing in power-efficient hardware.
Each hardware vendor will be able to provide tools and recommendations
for how to monitor power and cooling.
How to pick hardware for your Hadoop cluster
The first step in choosing a machine configuration is to understand
the type of hardware your operations team already manages. Operations
teams often have opinions about new machine purchases and will prefer to
work with hardware that they’re already familiar with. Hadoop is not
the only system that benefits from efficiencies of scale. Remember to
plan on using balanced hardware for an initial cluster when new to
Hadoop and if you do not yet understand your workload.
There are four types of nodes in a basic Hadoop cluster. We refer
here to a node as a machine performing a particular task. Most of the
machines will function as both datanodes and tasktrackers. As we
described, these nodes both store data and perform processing functions.
We recommend the following specifications for datanodes/tasktrackers in
a balanced Hadoop cluster:
- 4 1TB hard disks in a JBOD (Just a Bunch Of Disks) configuration
- 2 quad core CPUs, running at least 2-2.5GHz
- 16-24GBs of RAM (24-32GBs if you’re considering HBase)
- Gigabit Ethernet
The namenode is responsible for coordinating data storage on the
cluster and the jobtracker for coordinating data processing. The last
type of node is the secondarynamenode, which can be colocated on the
namenode machine for small clusters, and will run on the same hardware
as the namenode for larger clusters. We recommend our customers
purchase hardened machines for running the namenodes and jobtrackers,
with redundant power and enterprise-grade RAIDed disks. Namenodes also
require more RAM relative to the number of data blocks in the cluster. A
good rule of thumb is to assume 1GB of namenode memory for every one
million blocks stored in the distributed file system. With 100
datanodes in a cluster, 32GBs of RAM on the namenode provides plenty of
room to grow. We also recommend having a standby machine to replace the
namenode or jobtracker, in the case when one of these fails suddenly.
When you expect your Hadoop cluster to grow beyond 20 machines we
recommend that the initial cluster be configured as it were to span two
racks, where each rack has a top of rack gigabit switch, and those
switches are connected with a 10 GigE interconnect or core switch.
Having two logical racks gives the operations team a better understand
of the network requirements for inner-rack, and cross-rack
communication.
With a Hadoop cluster in place the team can start identifying
workloads and prepare to benchmark those workloads to identify CPU and
IO bottlenecks. After some time benchmarking and monitoring, the team
will have a good understanding as to how additional machines should be
configured. It is common to have heterogeneous Hadoop clusters
especially as they grow in size. Starting with a set of machines that
are not perfect for your workload will not be a waste.
Below is a list of various hardware configurations for different workloads, including our original “base” recommendation:
- Light Processing Configuration (1U/machine): Two quad core CPUs, 8GB
memory, and 4 disk drives (1TB or 2TB). Note that CPU-intensive work
such as natural language processing involves loading large models into
RAM before processing data and should be configured with 2GB RAM/core
instead of 1GB RAM/core.
- Balanced Compute Configuration (1U/machine): Two quad core CPUs, 16
to 24GB memory, and 4 disk drives (1TB or 2TB) directly attached using
the motherboard controller. These are often available as twins with two
motherboards and 8 drives in a single 2U cabinet.
- Storage Heavy Configuration (2U/machine): Two quad core CPUs, 16 to
24GB memory, and 12 disk drives (1TB or 2TB). The power consumption for
this type of machine starts around ~200W in idle state and can go as
high as ~350W when active.
- Compute Intensive Configuration (2U/machine): Two quad core CPUs,
48-72GB memory, and 8 disk drives (1TB or 2TB). These are often used
when a combination of large in-memory models and heavy reference data
caching is required.
Note that we expect to adopt 6 and 8 core configurations as they
arrive. The following diagram shows how a machine should be configured
according to workload:
Other hardware considerations
When we encounter applications that produce large amounts of
intermediate data–on the order of the same amount as is read in–we
recommend two ports on a single Ethernet card or two channel-bonded
Ethernet cards to provide 2 Gbps per machine. Alternatively for
customers who have already moved to 10 Gigabit Ethernet or Infiniband,
these solutions can be used to address network bound workloads. Be sure
that your operating system and BIOS are compatible if you’re
considering switching to 10 Gigabit Ethernet.
When computing memory requirements, factor in that Java uses up to
10% for managing the virtual machine. We recommend configuring Hadoop to
use strict heap size restrictions in order to avoid memory swapping to
disk. Swapping greatly impacts MapReduce job performance and can be
avoided by configuring machines with more RAM.
It is also important to optimize RAM for the memory channel width.
For example, when using dual-channel memory each machine should be
configured with pairs of DIMMs. With triple-channel memory each machine
should have triplets of DIMMs. This means a machine might end up with
18GBs (9x2GB) of RAM instead of 16GBs (4x4GB).
Conclusions
Purchasing appropriate hardware for a Hadoop cluster requires
benchmarking and careful planning to fully understand the workload.
However, Hadoop clusters are commonly heterogeneous and we recommend
deploying initial hardware with balanced specifications when getting
started.
