Hadoop Benchmark
Hadoop Benchmark
Hadoop Benchmark
Evaluating Cloudera,
Hortonworks, and MapR
in Micro-benchmarks and
Real-world Applications
Vladimir Starostenkov, Senior R&D Developer,
Kirill Grigorchuk, Head of R&D Department
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be
reproduced or transmitted in any form or by any means without written permission from the author.
Table of Contents
1. Introduction ........................................................................................................................................... 4
2. Tools, Libraries, and Methods ............................................................................................................... 5
2.1 Micro benchmarks ..................................................................................................................................................................6
2.1.1 WordCount .................................................................................................................................................................6
2.1.2 Sort ................................................................................................................................................................................7
2.1.3 TeraSort .......................................................................................................................................................................7
2.1.4 Distributed File System I/O ...................................................................................................................................7
2.2 Real-world applications ........................................................................................................................................................7
2.2.1 PageRank ....................................................................................................................................................................8
2.2.2 Bayes ............................................................................................................................................................................8
3. What Makes This Research Unique? ...................................................................................................... 9
3.1 Testing environment .............................................................................................................................................................9
4. Results .................................................................................................................................................. 11
4.1 Overall cluster performance ............................................................................................................................................. 11
4.2 Hortonworks Data Platform (HDP) ................................................................................................................................. 12
.................................................................................... 14
4.4 MapR ........................................................................................................................................................................................ 15
5. Conclusion ........................................................................................................................................... 18
Appendix A: Main Features and Their Comparison Across Distributions .............................................. 19
Appendix B: Overview of the Distributions ............................................................................................ 21
1. MapR ........................................................................................................................................................................................... 21
2. Cloudera .................................................................................................................................................................................... 22
3. Hortonworks ............................................................................................................................................................................ 23
Appendix C: Performance Results for Each Benchmarking Test ............................................................ 24
1. Real-world applications........................................................................................................................................................ 24
1.1 Bayes ............................................................................................................................................................................. 24
1.2 PageRank ..................................................................................................................................................................... 25
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
1. Introduction
Perhaps, there is hardly any expert in the big data field who has heard nothing about
Hadoop. Furthermore, very often Hadoop is used as a synonym to the term big data. It is
most likely, that the wide usage and popularity of this framework may have given a jumpstart to development of various distributions that derived from the initial open-source
edition.
The MapReduce paradigm was firstly introduced by Google and Yahoo continued with
development of Hadoop, which is based on this data processing method. From that moment
on, Hadoop has grown into several major distributions and dozens of sub-projects used by
thousands of companies. However, the rapid development of the Hadoop ecosystem and
the extension of its application area, lead to a misconception that Hadoop can be used to
solve any high-load computing task easily, which is not exactly true.
Actually, when a company is considering Hadoop to address its needs, it has to answer two
questions:
To collect information on these two points, companies spend an enormous amount of time
researching into distributed computing paradigms and projects, data formats and their
optimization methods, etc. This benchmark demonstrates performance results of the most
popular open-source Hadoop distributions, such as Cloudera, MapR, and Hortonworks. It
also provides all the information you may need to evaluate these options.
In this research, such solutions as Amazon Elastic MapReduce (Amazon EMR), Windows
Azure HDInsight, etc., are not analyzed, since they require uploading business data to public
clouds. This benchmark evaluates only stand-alone distributions that can be installed in
private data centers.
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
by Tom
To show how the amount of data changes on each stage of a MapReduce job, the whole
amount of input data was taken as 1.00. All the other indicators were calculated as ratios to
the input amount. For instance, the input data set was 100 GB (1.00) in size. After a Map task
had been completed, it increased to 142 GB (1.42), see Table 1. Using ratios instead of the
real data amounts allows for analyzing trends. In addition, these results can help to predict
the behavior of a cluster that deals with input data of a different size.
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
Map input
Combiner input
Combiner output
Reduce output
1.00
1.42
0.07
0.03
2.1.2 Sort
This workload sorts out unsorted text data; the amount of data on all the stages of the
Hadoop MapReduce process is the same. Being mostly I/O-bound, this workload has
moderate CPU utilization, as well as heavy disk and network I/O utilization (during the
shuffle stage). RandomTextWriter generates the input data.
Map input
Map output
Reduce output
1.0
1.0 (uncompressed)
1.0
2.1.3 TeraSort
TeraSort input data consists of 100-byte rows generated by the TeraGen application. Even
though this workload has high/moderate CPU utilization during Map/Reduce stages
respectively, it is mostly an I/O-bound workload.
Map input
Map output
Reduce output
1.0
0.2 (compressed)
1.0
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
2.2.1 PageRank
PageRank is a widely-known algorithm that evaluates and ranks Web sites in search results.
To calculate a rating, a PageRank job is repeated several times, which is an iterative CPUbound workload. The benchmark consists of several chained Hadoop jobs, each represented
by a separate row in the table below. In this benchmark, PageRank had two HDFS blocks per
CPU core, which is the smallest input per node in this test.
Map input
Combiner input
Combiner output
Reduce output
1.0
1.0E-005
1.0E-007
1.0E-008
1.0
5.0
1.0
1.0
0.1
0.1
2.2.2 Bayes
The next application is a part of the Apache Mahout project. The Bayes Classification
workload has rather complex patterns of accessing CPU, memory, disk, and network. This
test creates a heavy load on a CPU when completing Map tasks. However in this case, this
workload hit an I/O bottleneck.
Bayes
+1 650
Map input
Combiner input
Combiner output
Reduce output
1.0
28.9
22.1
19.4
19.4
14.4
12.7
7.4
7.4
9.3
4.8
4.6
7.4
3.1
1.0E-004
1.0E-005
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
In general, virtualized cloud environments provide flexibility in tuning, which was required
to carry out tests on clusters of a different size. In addition, cloud infrastructure allows for
obtaining more unbiased results, since all tests can be easily repeated to verify their results.
For this benchmark, all tests were run on the ProfitBricks virtualized infrastructure. The
deployment settings selected for each distribution provided similar test conditions, as much
as it was possible.
In this research, we tested distributions that include updates, patches, and additional
features that ensure stability of the framework. Hortonworks and Cloudera are active
contributors to Apache Hadoop and they provide fast bug fixing for their solutions.
Therefore, their distributions are considered more stable and up-to-date.
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
10
The service provides a wide range of configuration parameters for each node, for instance,
the CPU capacities of a node can vary from 1 to 48 cores, and RAM can be from 1 to 196 GB
per node. This variety of options allows for achieving the optimal CPU/RAM/network/storage
balance for each Hadoop task.
Unlike Amazon that offers preconfigured nodes, ProfitBricks allows for manual tuning of
each node based on your previous experience and performance needs. InfiniBand is a
modern technology used by ProfitBricks. It allowed for achieving the maximum inter-node
communication performance inside a data center.
Cluster configuration:
Each node had four CPU cores, 16 GB of RAM, and 100 GB of virtualized disk space. Cluster
size ranged from 4 to 16 nodes. Nodes required for running Ganglia and cluster
management were not included into this configuration. The top cluster configuration
featured 64 computing cores and 256 GB of RAM for processing 1.6 TB of test data.
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
11
4. Results
Each Cloudera and Hortonworks DataNode contained one disk. MapR distribution was
evaluated in a slightly different way. Three data disks were attached to each DataNode
following MapR recommendations. Taking this into account, MapR was expected to perform
I/O sensitive tasks three times faster. However, the actual results were affected by some
peculiarities of virtualization (see Figure 11).
The comparison of Cloudera and Hortonworks features showed that these two distributions
are very similar (see Appendix A). It was also proved by the results of the tests (see Appendix
B). The overall performance of Hortonworks and Cloudera clusters is demonstrated by Figure
4 and 6 respectively.
MapR
Overall cluster performance
5
4
3
2
1
0
BAYES
DFSIOE
HIVEAGGR
PAGERANK
4
12
SORT
TERASORT WORDCOUNT
16
Figure 2. The overall performance results of the MapR distribution in all benchmark tests
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
12
MapR
Performance per Node
1.2
1
0.8
0.6
0.4
0.2
0
BAYES
DFSIOE
HIVEAGGR
PAGERANK
4
12
SORT
TERASORT
WORDCOUNT
16
Figure 3. The average performance of a single node of the MapR cluster in all benchmark tests
Cluster performance scales linearly under the WordCount workload. It behaves the same in
running PageRank until the cluster reaches an I/O bottleneck. The results of other
benchmarks strongly correlated with DFSIO. Disk I/O throughput did not scale in this test
environment, however, analyzing the reasons for that was not the focus of this research. To
learn more about the drawbacks of Hadoop virtualization, read
Sammer, C
As it was mentioned before, MapR had three disks per node. In case all nodes are hosted on
the same server, the virtual cluster utilizes disk bandwidth which is obviously limited
much faster. ProfitBricks allows for hosting up to 48-62 cores on the same server, which is
equivalent to a cluster that consists of 12 15 nodes with the configuration described in this
benchmark (four cores per node).
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
13
Hortonworks
Overall cluster performance
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
BAYES
DFSIOE
HIVEAGGR
PAGERANK
4
12
SORT
TERASORT
WORDCOUNT
16
Figure 4. The overall performance results of the Hortonworks cluster in all benchmark tests
Hortonworks
Performance per Node
1.2
1
0.8
0.6
0.4
0.2
0
BAYES
DFSIOE
HIVEAGGR
PAGERANK
4
12
SORT
TERASORT
WORDCOUNT
16
Figure 5. The average performance of a single node of the Hortonworks cluster in all benchmark
tests
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
14
(CDH)
The Cloudera Hadoop distribution showed almost the same performance as Hortonworks,
except for Hive queries, where it was slower.
Cloudera
Overall cluster performance
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
BAYES
DFSIOE
HIVEAGGR
4
PAGERANK
8
12
SORT
TERASORT WORDCOUNT
16
Figure 6. The overall performance results of the Cloudera cluster in all benchmark tests
Cloudera
Performance per Node
1.2
1
0.8
0.6
0.4
0.2
0
12
16
Figure 7. The average performance of a single node of the Cloudera cluster in all benchmark tests
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
15
WordCount
WordCount
1.1
1.05
MapR
MapR
1
Hortonworks
Hortonworks
0.95
Cloudera
Cloudera
0.9
4
12
16
12
16
4.4 MapR
The performance results of the MapR cluster under the Sort load were quite unexpected.
Number of Nodes
MapR
Number of Nodes
MapR
Baseline
0.5
1.5
Baseline
0.5
1.5
16
MapR
Number of Nodes
Baseline
0.2
0.4
0.6
0.8
1.2
1.4
16
12
MapR
Number of Nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 10. The MapR performance results in the DFSIO (write) benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
17
The virtualized disks of a 4-node cluster showed the reading/writing speed of 250/700
MB/sec. The overall cluster performance grew not in a linear way (see Figure 11), meaning
that the total speed of data processing can be improved by the optimal combination of CPU,
RAM, and disk space parameters.
DFSIO Read
Overall cluster performance
1.7
1.6
1.5
1.4
1.3
MapR
1.2
Hortonworks
1.1
Cloudera
1
0.9
0.8
4
12
16
DFSIO Write
Overall cluster performance
1.4
1.2
1
MapR
0.8
Hortonworks
0.6
Cloudera
0.4
0.2
4
12
16
Figure 11. Performance results for the MapR, Hortonworks, and Cloudera distributions in the
DFSIO (read/write) benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
18
5. Conclusion
Despite of the fact that the configuration of a cluster deployed in the cloud was similar to
that of the one deployed on bare metal, the performance and scalability of the virtualized
solution were different. In general, Hadoop deployed on bare metal is expected to scale
linearly, until inter-node communication will start to slow it down or it reaches the limits of
the HDFS, which is around several thousand of nodes.
The actual measurements showed that even though the overall performance was very high,
it was affected by the limited total disk throughput. Therefore, the disk I/O became a serious
bottleneck whereas the computing capacities of the cluster were not fully utilized. Apache
Spark, which was announced by Cloudera when this research was conducted, or GridGain
In-Memory Accelerator for Hadoop can be suggested for using in the ProfitBricks
environment.
It can be assumed that the type of Hadoop distribution has a much less considerable impact
on the overall system throughput than the configuration of the MapReduce task parameters.
For instance, the TeraSort workload was processed 2 3 times faster when the parameters
described in Appendix E were tuned specifically for this load. By configuring these settings,
you can achieve 100% utilization of your CPU, RAM, disk, and network. So, the performance
of each distribution can be greatly improved by selecting proper parameters for each
specific load.
Running Hadoop in clouds allows for fast horizontal and vertical scaling, however, there are
fewer possibilities for tuning each part of the infrastructure. In case you opt for a virtualized
deployment, you should select a hosting/IaaS provider that gives freedom in configuring
your infrastructure. To achieve optimal utilization of resources, you will need information on
the parameters set for network and disk storage and have a possibility to change them.
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
19
Implementation Hortonworks
HDP -1.3
File system
HDFS 1.2.0
HDFS 2.0.0
- non-Hadoop access
NFSv3
WebHDFS
HttpFS
1.2.0
0.20.2
**
- Web access
MapReduce
MapR-FS
Cascading
2.1
Non-relational database
Apache HBase
0.94.6.1
0.94.6
0.92.2
Metadata services
0.5.0
0.4.0
Scripting platform
Apache Pig
0.11
0.11.0
0.10.0
DataFu
0.0.4
Apache Hive
0.11.0
0.10.0
0.9.0
Workflow scheduler
Apache Oozie
3.3.2
3.3.2
3.2.0
Cluster coordination
Apache
Zookeeper
3.4.5
3.4.5
3.4(?)
Apache Sqoop
1.4.3
1.4.3
1.4.2
1.3.1
1.3.0
1.2.0
0.7.0
0.7
0.7
+1 650 395-7002
Mahout
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
20
Hadoop UI
Hue
2.2.0
2.3.0
Talend Open
Studio for Big
Data
5.3
Cloud services
Whirr
0.8.2
0.7.0
Tez (Stinger)
Impala
****
Search
0.1.5
Administration
MapR Control
System
- installation
- monitoring
Ganglia
Nagios
- fine-grained authorization
Sentry
Splitting resource
management and scheduling
YARN
1.1
2.0.4
2.0.0
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
21
Summary
MapR has the MapRFS feature which is a substitute for the standard HDFS. Unlike HDFS, this
system aims to sustain deployments that consist of up to 10,000 of nodes with no single
point of failure, which is guaranteed by the distributed NameNode. MapR allows for storing
1 10 Exabytes of data and provides support for NFS and random read-write semantics. It is
stated by the MapR developers that elimination of the Hadoop abstraction layers can help to
increase performance 2x.
There are three editions of the MapR distribution: M3, which is completely free, M5 and M7,
the latter two are paid enterprise versions. Although M3 provides unlimited scalability and
NFS, it does not ensure high availability and snapshots that are available in M5 or instant
recovery of M7. M7 is an enterprise-level platform for NoSQL and Hadoop deployments.
MapR distribution is available as a part of Amazon Elastic MapReduce and Google Cloud
Platform.
Notable customers and partners
MapR M3 and M5 editions are available as premium options for Amazon Elastic MapReduce;
Google partnered with MapR in launching Compute Engine;
Cisco Systems announced support for MapR software on the UCS platform;
comScore
The company
Based in San Jose, California, MapR focuses on development of Hadoop-based projects. The
company contributes to such projects as HBase, Pig, Apache Hive, and Apache ZooKeeper.
After signing an agreement with EMC in 2011, the company supplies a specific Hadoop
engineering@altoros.com
+1 650 395-7002
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
22
distribution tuned for EMC hardware. MapR also partners with Google and Amazon in
improving their Elastic Map Reduce (EMR) service.
2. Cloudera
Summary
Of all the distributions analyzed in this research,
s solution has the most powerful
Hadoop deployment and administration tools designed for managing a cluster of an
unlimited size. It is also open-source and the company is an active contributor to Apache
Hadoop. Cloudera is a major Apache Hadoop contributor. In addition, the Cloudera
distribution has its own native components, such as Impala, a query engine for massive
parallel processing and Cloudera Search powered by Apache Solr.
Notable customers and partners
eBay
CBS Interactive
Qualcomm
Expedia
The company
Based in Palo Alto, Cloudera is one of the leading companies that provides Hadoop-related
services and trainings for the staff
statistics, more than 50% of
its efforts are dedicated to improving such open-source projects as Apache Hive, Apache
Avro, Apache HBase, etc. that are a part of a large Hadoop ecosystem. In addition, Cloudera
invests into Apache Software Foundation, a community of developers who contribute to the
family of Apache software projects.
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
23
3. Hortonworks
Summary
Being 100% open-source, Hortonworks is strongly committed to Apache Hadoop and it is
one of the main contributors to the solution.
brings high performance, scalability, and SQL compliance to Hadoop deployments. YARN,
the Hadoop OS, and Apache Tez, a framework for near real-time big data processing, help
Stringer to speed up Hive and Pig by up to 100x.
As a result of Hortonworks partnership with Microsoft, HDP is the only Hadoop distribution
available as a native component of Windows Server. A Windows-based Hadoop cluster can
be easily deployed on Windows Azure through HDInsight Service.
Notable customers and partners
Western Digital
eBay
Samsung Electronics
The company
Hortonworks is a company headquartered in Palo Alto, California. Being a sponsor of the
Apache Software Foundation and one of the main contributors to Apache Hadoop, the
company specializes in providing support for Apache Hadoop. The Hortonworks distribution
includes such components as HDFS, MapReduce, Pig, Hive, HBase, and Zookeeper. Together
with Yahoo!, Hortonworks hosts the annual Hadoop Summit event, the leading conference
for the Apache Hadoop community.
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
24
Bayes
Overall cluster performance
1.45
1.4
1.35
1.3
MapR
1.25
Hortonworks
1.2
Cloudera
1.15
1.1
1.05
1
4
12
16
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
25
Bayes
Performance per Node
1.2
1
0.8
MapR
0.6
Hortonworks
Cloudera
0.4
0.2
0
4
12
16
Figure 13: Bayes: the performance of a single node for each cluster size
1.2 PageRank
PageRank
Overall cluster performance
4.5
4
3.5
MapR
Hortonworks
2.5
Cloudera
2
1.5
1
12
16
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
26
PageRank
Performance per Node
1.05
1
0.95
0.9
MapR
Hortonworks
0.85
Cloudera
0.8
0.75
0.7
4
12
16
Figure 15. PageRank: the performance of a single node for each cluster size
2. Micro benchmarks
2.1 Distributed File System I/O (DFSIO)
DFSIO
Overall cluster performance
1.8
1.7
1.6
1.5
MapR
1.4
Hortonworks
1.3
Cloudera
1.2
1.1
1
12
16
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
27
DFSIO
Performance per Node
1.2
1
0.8
MapR
0.6
Hortonworks
Cloudera
0.4
0.2
0
4
12
16
Figure 17. DFSIO: the performance of a single node for each cluster size
HIVEAGGR
Overall cluster performance
2.6
2.4
2.2
2
MapR
1.8
Hortonworks
1.6
Cloudera
1.4
1.2
1
4
12
16
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
28
HIVEAGGR
Performance per Node
1.2
1
0.8
MapR
0.6
Hortonworks
Cloudera
0.4
0.2
0
4
12
16
Figure 19. Hive aggregation: the performance of a single node for each cluster size
2.3 Sort
Sort
Overall cluster performance
2
1.9
1.8
1.7
1.6
MapR
1.5
Hortonworks
1.4
Cloudera
1.3
1.2
1.1
1
4
12
16
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
29
Sort
Performance per Node
1.2
1
0.8
MapR
0.6
Hortonworks
Cloudera
0.4
0.2
0
4
12
16
Figure 21. Sort: the performance of a single node for each cluster size
2.4 TeraSort
TeraSort
Overall cluster performance
2.8
2.6
2.4
2.2
MapR
2
1.8
Hortonworks
1.6
Cloudera
1.4
1.2
1
4
+1 650 395-7002
12
16
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
30
TeraSort
Performance per Node
1.2
1
0.8
MapR
0.6
Hortonworks
Cloudera
0.4
0.2
0
4
12
16
Figure 23. TeraSort: the performance of a single node for each cluster size
2.5 WordCount
WordCount
The overall cluster performance
4.5
4
3.5
MapR
Hortonworks
2.5
2
Cloudera
1.5
1
4
12
16
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
31
WordCount
Performance per Node
1.06
Performance ratios
1.04
1.02
1
MapR
0.98
Hortonworks
0.96
Cloudera
0.94
0.92
0.9
4
12
16
Figure 25. WordCount: the performance of a single node for each cluster size
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
32
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
1.4
1.6
16
12
Figure 26. The overall performance of the MapR cluster in the Bayes benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
33
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 27. The performance of a single node of the MapR cluster in the Bayes benchmark
MapR
Number of nodes
Baseline
0.5
1.5
16
+1 650 395-7002
12
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
34
Figure 28. The overall performance of the MapR cluster in the DFSIO (read) benchmark
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 29. The performance of a single node of the MapR cluster in the DFSIO (read) benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
35
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
1.4
16
12
Figure 30. The overall performance of the MapR cluster in the DFSIO (write) benchmark
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
+1 650 395-7002
12
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
36
Figure 31. The performance of a single node of the MapR cluster in the DFSIO (write) benchmark
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
1.4
16
12
Figure 32. The overall performance of the MapR cluster in the DFSIO benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
37
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 33. The performance of a single node of the MapR cluster in the DFSIO benchmark
MapR
Number of nodes
Baseline
0.5
1.5
16
12
Figure 34. The overall performance of the MapR cluster in the Hive aggregation benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
38
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 35. The performance of a single node of the MapR cluster in the Hive aggregation
benchmark
MapR
Number of nodes
Baseline
0.5
1.5
2.5
3.5
12
Figure 36. The overall performance of the MapR cluster in the PageRank benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
39
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 37. The performance of a single node of the MapR cluster in the PageRank benchmark
MapR
Number of nodes
Baseline
0.5
1.5
12
Figure 38. The overall performance of the MapR cluster in the Sort benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
40
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 39. The performance of a single node of the MapR cluster in the Sort benchmark
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
1.4
16
12
Figure 40. The overall performance of the MapR cluster in the TeraSort benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
41
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 41. The performance of a single node of the MapR cluster in the TeraSort benchmark
MapR
Number of nodes
Baseline
12
Figure 42. The overall performance of the MapR cluster in the WordCount benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
42
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 43. The performance of a single node of the MapR cluster in the WordCount benchmark
2. Hortonworks
Hortonworks
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
1.4
+1 650 395-7002
12
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
43
Figure 44. The overall performance of the Hortonworks cluster in the Bayes benchmark
Hortonworks
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
12
Figure 45. The performance of a single node of the Hortonworks cluster in the Bayes benchmark
Hortonworks
Number of nodes
Baseline
0.5
1.5
16
+1 650 395-7002
12
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
44
Figure 46. The overall performance of the Hortonworks cluster in the DFSIO (read) benchmark
Hortonworks
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 47. The performance of a single node of the Hortonworks cluster in the DFSIO (read)
benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
45
Hortonworks
Number of nodes
Baseline
0.5
1.5
16
12
Figure 48. The overall performance of the Hortonworks cluster in the DFSIO (write) benchmark
Hortonworks
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
+1 650 395-7002
12
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
46
Figure 49. The performance of a single node of the Hortonworks cluster in the DFSIO (write)
benchmark
Hortonworks
Number of nodes
Baseline
0.5
1.5
16
12
Figure 50. The overall performance of the Hortonworks cluster in the DFSIO benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
47
Hortonworks
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 51. The performance of a single node of the Hortonworks cluster in the DFSIO benchmark
Hortonworks
Number of nodes
Baseline
0.5
1.5
2.5
12
Figure 52. The overall performance of the Hortonworks cluster in the Hive aggregation
benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
48
MapR
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
12
Figure 53. The performance of a single node of the Hortonworks cluster in the Hive aggregation
benchmark
Hortonworks
Number of nodes
Baseline
12
Figure 54. The overall performance of the Hortonworks cluster in the PageRank benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
49
Hortonworks
Number of nodes
Baseline
0.9
0.92
0.94
0.96
0.98
1.02
16
12
Figure 55. The performance of a single node of the Hortonworks cluster in the PageRank
benchmark
Hortonworks
Number of Nodes
Baseline
0.5
1.5
12
Figure 56. The overall performance of the Hortonworks cluster in the Sort benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
50
Hortonworks
Number of Nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 57. The performance of a single node of the Hortonworks cluster in the Sort benchmark
Hortonworks
Number of Nodes
Baseline
0.5
1.5
2.5
12
Figure 58. The overall performance of the Hortonworks cluster in the TeraSort benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
51
Hortonworks
Number of Nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 59. The performance of a single node of the Hortonworks cluster in the TeraSort
benchmark
Hortonworks
Number of nodes
Baseline
12
Figure 60. The overall performance of the Hortonworks cluster in the WordCount benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
52
Hortonworks
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
12
Figure 61. The performance of a single node of the Hortonworks cluster in the WordCount
benchmark
3. Cloudera
Cloudera
Numbers of nodes
Baseline
0.5
1.5
+1 650 395-7002
12
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
53
Figure 62. The overall performance of the Cloudera cluster in the Bayes benchmark
Cloudera
Numbers of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
12
Figure 63. The performance of a single node of the Cloudera cluster in the Bayes benchmark
Cloudera
Number of nodes
Baseline
0.5
1.5
12
Figure 64. The overall performance of the Cloudera cluster in the DFSIO (read) benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
54
Cloudera
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
12
Figure 65. The performance of a single node of the Cloudera cluster in the DFSIO (read)
benchmark
Cloudera
Number of nodes
Baseline
0.5
1.5
16
12
Figure 66. The overall performance of the Cloudera cluster in the DFSIO (write) benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
55
Cloudera
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 67. The performance of a single node of the Cloudera cluster in the DFSIO (write)
benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
56
Cloudera
Number of nodes
Baseline
0.5
1.5
16
12
Figure 68. The overall performance of the Cloudera cluster in the DFSIO benchmark
Cloudera
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
16
12
Figure 69. The performance of a single node of the Cloudera cluster in the DFSIO benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
57
Cloudera
Number of nodes
Baseline
0.5
1.5
2.5
12
Figure 70. The overall performance of the Cloudera cluster in the Hive aggregation benchmark
Cloudera
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
12
Figure 71. The performance of a single node of the Cloudera cluster in the Hive aggregation
benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
58
Cloudera
Number of nodes
Baseline
16
12
Figure 72. The overall performance of the Cloudera cluster in the PageRank benchmark
Cloudera
Number of nodes
0.9
Baseline
0.92
0.94
0.96
0.98
1.02
12
Figure 73. The performance of a single node of the Cloudera cluster in the PageRank benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
59
Cloudera
Number of nodes
Baseline
0.5
1.5
16
12
Figure 74. The overall performance of the Cloudera cluster in the Sort benchmark
Cloudera
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
12
Figure 75. The performance of a single node of the Cloudera cluster in the Sort benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
60
Cloudera
Number of nodes
Baseline
0.5
1.5
2.5
16
12
Figure 76. The overall performance of the Cloudera cluster in the TeraSort benchmark
Cloudera
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
12
Figure 77. The performance of a single node of the Cloudera cluster in the TeraSort benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
61
Cloudera
Number of nodes
Baseline
16
12
Figure 78. The overall performance of the Cloudera cluster in the WordCount benchmark
Cloudera
Number of nodes
Baseline
0.2
0.4
0.6
0.8
1.2
12
Figure 79. The performance of a single node of the Cloudera cluster in the WordCount benchmark
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
62
1.5
1.4
1.3
MapR
1.2
Hortonworks
1.1
Cloudera
1
0.9
0.8
4
12
16
Figure 80. The overall performance of each distribution in the DFSIO-read benchmark, sectioned
by the cluster size
DFSIO Read
Performance per Node
1.1
0.9
MapR
0.7
Hortonworks
0.5
Cloudera
0.3
4
12
16
Figure 81. The performance of a single node of each distribution in the DFSIO-read benchmark,
sectioned by the cluster size
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
63
Hortonworks
0.6
Cloudera
0.4
0.2
4
12
16
Figure 82. The overall performance of each distribution in the DFSIO-write benchmark, sectioned
by the cluster size
DFSIO Write
Performance per Node
1.1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
MapR
Hortonworks
Cloudera
12
16
Figure 83. The performance of a single node of each distribution in the DFSIO-write benchmark,
sectioned by the cluster size
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
64
Description
mapred.map.tasks
mapred.reduce.tasks
mapred.output.compress
Set true in order to compress the output of the MapReduce job, yse
mapred.output.compression.codec to specify the compression
codec.
mapred.map.child.java.opts
io.sort.mb
A Map task output buffer size. Use this value to control the spill
process. When buffer is filled up to a io.sort.spill.percent a
background spill thread is started.
mapred.job.reduce.input.buffer.percent
mapred.inmem.merge.threshold
mapred.job.shuffle.merge.percent
The threshold usage proportion for the Map outputs buffer for
starting the process of merging the outputs and spilling to disk
mapred.reduce.slowstart.completed.m
aps
dfs.replication
dfs.block.size
mapred.task.timeout
mapred.map.tasks.speculative.executio
n
mapred.job.reuse.jvm.num.tasks
The maximum number of tasks to run for a given job for each JVM
io.sort.record.percent
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.
65
of the Map outputs. The remaining space is used for the Map
output records themselves.
Example
Table 7. Parameters that were tuned to achieve optimal performance of Hadoop jobs
About the author:
Vladimir Starostenkov is a Senior R&D Engineer at Altoros, a company that focuses on
accelerating big data projects and platform-as-a-service enablement. He has more than five
years of experience in implementing complex software architectures, including dataintensive systems and Hadoop-driven applications. Having strong background in physics
and computer science, Vladimir is interested in artificial intelligence and machine learning
algorithms.
About Altoros:
Altoros is a big data and Platform-as-a-Service specialist that provides system integration for
IaaS/cloud providers, software companies, and information-driven enterprises. The company
builds solutions on the intersection of Hadoop, NoSQL, Cloud Foundry PaaS, and multi-cloud
deployment automation. For more, please visit www.altoros.com or follow @altoros.
Liked this white paper?
Share it on the Web!
+1 650 395-7002
engineering@altoros.com
www.altoros.com | twitter.com/altoros
2013 Altoros Systems, Inc. Any unauthorized republishing, rewriting or use of this material is prohibited. No part of this resource may be reproduced or transmitted in
any form or by any means without written permission from the author.