K means clustering in the cloud - a mahout test

2011 Workshops of International Conference on Advanced Information Networking and Applications

K-means clustering in the cloud - a Mahout test
Rui Máximo Esteves * Chunming Rong*
rui.esteves@uis.no chunming.rong@uis.no

Rui Pais * * Department of Electrical and Computer Engineering,
rma.pais@stud.uis.no University of Stavanger, Norway

Abstract—. The K-Means is a well known clustering algorithm
that has been successfully applied to a wide variety of problems. II. THEORY
However, its application has usually been restricted to small
datasets. Mahout is a cloud computing approach to K-Means that Clustering
runs on a Hadoop system. Both Mahout and Hadoop are free
and open source. Due to their inexpensive and scalable
Clustering refers to the process of grouping samples into
characteristics, these platforms can be a promising technology to
solve data intensive problems which were not trivial in the past. different classes based on their similarities. Samples within a
In this work we studied the performance of Mahout using a large class have high similarity in comparison to one another but
data set. The tests were running on Amazon EC2 instances and are very dissimilar to samples in other classes. These groups
allowed to compare the gain in runtime when running on a multi or classes are called clusters. Clustering algorithms can
node cluster. This paper presents some results of ongoing divide the samples into clusters automatically without any
research. preconceptions about what kind of groupings should be
found. Within the context of machine learning, clustering is
considered to be a form of unsupervised learning since there
is no target variable to guide the learning process. Unlike
Keywords- K-means, cloud computing, mahout, map reduce classification, clustering and unsupervised learning do not
rely on predefined classes and class-labeled training
I. INTRODUCTION examples. For this reason, clustering is a form of learning by
observation, rather than learning by examples [2]. The
The K-means is one of the most frequently used clustering number of clusters, size, and shape are not in general known
algorithms. It is simple and straightforward and has been in anticipation, and each of these parameters must be
successfully applied during the last few years. Under the determined by either the user or the clustering algorithm [3].
assumption that datasets tend to be small, research on Clustering has been used in many areas, including data
clustering algorithms has traditionally focused on improving mining, statistics, biology, and machine learning. Clustering
the quality of clustering. However, many datasets now are can also be used for outlier detection, where outliers (values
large and cannot fit into main memory [1]. that are “far away” from any cluster) may be more
Mahout/Hadoop can be a promising and inexpensive interesting than common cases [2]. One useful application
solution to solve problems with large data sets. However, example is fault detection [4, 5]. Other examples of
there is a lack of studies regarding its performance. With this applications can be found in [2, 6]. The concept of clustering
work we wanted to study the gain in runtime when using this is not particularly new, but it is still an important topic of
solution and also to test if there was any loss in the clusters research. A useful survey about recent advances can be
quality. To meet those aims we used Amazon EC2 instances found in [7].
and set up a small cluster. We chose to study K-means since
it is one of the most popular and easy algorithms. Cluster algorithms types
The rest of this paper is organized in the following way:
section II presents some theoretical foundations. The The two major types of cluster algorithms are
experiments are described in section III and section IV the hierarchical and partitional.
results, interpretation and discussion are presented. Future The first type produces a hierarchical decomposition of
work is in section V and the final conclusions are in section the data set into partitions. It merges or splits the partitions
VI. into clusters using a similarity criterion.
In the second type, an algorithm creates an initial division
of the data set, according to a given number of partitions. It
then uses an iterative relocation technique that attempts to
improve the partitioning by moving objects from one group
to another. The general criterion for partitioning is a
combination of high similarity of the samples inside of
clusters with high dissimilarity between distinct clusters.
This means that samples in the same cluster are “close” or

978-0-7695-4338-3/11 $26.00 © 2011 IEEE 514
DOI 10.1109/WAINA.2011.136

related to each other, whereas samples of different clusters the input to be processed by the reduce tasks [14]. Hadoop
are “far apart” or very different. [2, 6] uses a distributed file system HDFS, which creates multiple
Each clustering type is better suited than to a particular replicas of data blocks and distributes them on computer
type of problem. Abbas [8] concluded that partitional nodes throughout a cluster to enable reliable, extremely rapid
algorithms are suited for large data set while hierarchical are computations [15].
for small data sets. Singla, Yadav and Sing [9] presented Mahout3 is a machine learning library that runs over a
some arguments in favor of partitional algorithm. According Hadoop system. It has a collection of algorithms to solve
to them, there is no need to determine the full dendrogram clustering, classification and prediction problems. It uses
with a partitional algorithm. They also found that the time MapReduce paradigm which in combination with Hadoop
complexity of partitonal algorithm is little compared with can be used as an inexpensive solution to solve machine
hierarchical. Zhao and Kaprypis [10] had similar results. On learning problems.
the other end, hierarchical algorithms have the advantages of Mahout is a project still in development which has been
an easily interpreted graphical representation and ignorance used mainly for recommendation engines, document
of influence from initialization and local minima [11]. classifiers and to solve other web typical problems. At the
moment, its usage for clustering is not sufficiently explored.
K-means Hadoop and Mahout are free and open source projects.
Due to their inexpensive and scalable characteristics, these
K-means clustering is a widely used partition algorithm. platforms can be a promising technology to solve data
It partitions the samples into clusters by minimizing a intensive problems which were not trivial in the past.
measure between the samples and the centroid of the However, it is important to study the tradeoff between
clusters. Euclidean distance is a similarity criterion the overhead of using Map/Reduce and the gain of
extensively used for samples in Euclidean space. Manhattan performance by distributing the computation. It is also
distance is a simpler measure also for Euclidean space. essential to check if the clusters maintain their quality.
Cosine distance and Jaccard is often employed for Mahout contains various implementations of clustering,
documents clustering. [12] like K-means, fuzzy K-means, meanshift and Dirichlet
The algorithm for the standard K-means is given as among others. To input the data for Mahout clustering it is
follows [13]: necessary to do some procedures first. If the data is not
1. Choose a number of clusters k numerical it has to be first preprocessed. It is required then to
2. Initialize centroid1 c1,… ck create vectors. If the data set is sparse it allows the user to
3. For each data point, compute the centroid it is closest create sparse vectors that are much more compact. The
to (using some distance measure) and assign the data vectors are finally converted to a specific Hadoop file format
point to this cluster. that is SequenceFile.
4. Re-compute centroids (mean of data points in The K-means clustering algorithm takes the following
cluster) input parameters:
5. Stop when there are no new re-assignments. - A SequenceFile containing the input vectors.
- A SequenceFile containing the initial cluster centers. If
The K-means clustering is simple but it has high time not present it will attribute them randomly.
complexity when the data sets are large. In these - A similarity measure to be used.
circumstances the memory of a single machine can be a - The convergence Threshold that will be the stopping
restriction. As a consequence it has not been used in the past condition of the K-means. If in a particular iteration, any
with large data sets [13]. centers of the clusters do not change beyond that threshold,
then no further iterations are done.
Hadoop/Mahout solution - The maximum number of iterations to be processed if it
does not converge first.
Hadoop 2 is a software framework which allows the - The number of reducers to be used. This value
effortless development of cloud computing systems. It determines the parallelism of the execution.
supports data intensive distributed applications on large - The vector implementation used for the input files.
clusters built of commodity hardware. This framework is As output the user gets the centroids coordinates and the
designed to be scalable, which allows the user to add more samples attributed to each cluster. The output files are in
nodes as the application requires. Hadoop uses a distributed SequenceFile format. Mahout provides the necessary tools
computing paradigm named MapReduce. for file conversion and creating the vectors [16].
The MapReduce programming model consists of two There are three stages in a K-means job (figure 1):
user defined functions: map and reduce, specified on a job. - Initial stage: the segmentation of the dataset into
The job usually splits the input dataset into independent HDFS blocks; their replication and transfer to other
blocks which are processed by the map tasks in a parallel machines; according to the number of blocks and
way. Hadoop sorts the outputs of the maps, which are then cluster configuration it will be assigned and
distributed the necessary tasks.
1
Centroid refers to the center of the cluster.
2 3
http://hadoop.apache.org/ http://mahout.apache.org/

515

INITIAL STAGE MAP STAGE REDUCE STAGE
Dataset Block
avg

(…)

Centroides avg
Task
vector
destribution

Figure 1 – The three stages of a K-means job.

- Map stage: calculates the distances between samples
and centroids; match samples with the nearest III. EXPERIMENTS
centroid and assigns them to that specific cluster.
Each map task is processed with a data block. The dataset
Reduce stage: recalculates the centroid point using the
average of the coordinates of all the points in that cluster. To evaluate the performance of multi node versus single
The associated points are averaged out to produce the new node we used a dataset from the 1999 kdd cup4. This dataset
location of the centroid. The controids configuration is was gathered by Lincoln Labs: nine weeks of raw TCP dump
feedback into the Mappers. The loop ends when the data were collected from a typical US air force LAN. During
centroids converge. the use of the LAN, several attacks were performed on it.
After finishing the job Mahout provides the K-means The raw packet data were then aggregated into connection
time, the centroids and the clustered samples. data. Per record, extra features were derived, based on
domain knowledge about network attacks. There are 38
Quality of clustering different attack types, belonging to 4 main categories. It has
4,940,000 records, with 41 attributes and 1 label (converted
An important problem in clustering is how to decide what to numerical. A 1.1 GB dataset was used. This file was
the best set of clusters is for a given data set, in terms of both randomly segmented into smaller files. For the data
the number of clusters and the membership of those clusters preprocessing we developed a Java software.
[17].
One simple inter-cluster quality measure was used in this The Machines
work: the sum of the distances squared between centroids
(Q1) [18]. We ran our tests using Amazon EC2 m1.large instances
with the following characteristics: 7.5 GB memory; 4 EC2
Compute Units (2 virtual cores with 2 EC2 Compute Units
each); 850 GB of storage; 64-bit platform; high I/O
(eq. 1) Performance.
All the instances were launched using the same Amazon
where: ci is the centroid of a cluster ci; Machine Image (AMI). We adapted an existing Cloudera
d is the distance metric; AMI5 to our specific needs and stored on Amazon S3. Our
and k is the number of clusters. AMI uses Mahout 0.4 and Hadoop 0.20.2+237 and we added
scripts for automatic cluster configuration. This way the user
In our study, we do not aim to reach the optimal cluster just needs to launch EC2 instances with this AMI, run a
configuration. Instead, we want to check if the quality of the script on the master node and the cluster became ready for
Mahout K-means is maintained between the tests. As so, we use. The HDFS configuration used was the default one,
chosen this measure since it is not computationally where the block size is 64 MB and the replication factor 3.
expensive and easy to implement. Jobs
Related work To study how the performance changes with the
scalability, K-means jobs were setup to guaranty that the
At this date we could not find any intensive tests of algorithm will converge. K-means can produce different
Mahout in the literature. However there are two interesting results depending on the position of the initial centroids. To
works with scaling tools for data mining using MapReduce 4
[19, 20]. http://www.inf.ed.ac.uk/teaching/courses/dme/html/datasets0405.html
5
cloudera-hadoop-ubuntu-20090623_x86_64

516

avoid this issue, the same initial centroids were used for all
experiments. Each test was repeated 3 times to ensure the
results were not affected by Amazon fluctuations.
The Mahout K-means algorithm was executed using the
following parameters: Euclidean distance as the similarity
measure on experiments A and B; Squared Euclidean on C.
All experiments were set with 150 maximum interactions,
and 11 clusters.
To guarantee that the quality of the Mahout K-means was
maintained the equation 1 was applied.

IV. RESULTS AND INTERPRETATION

A – How does Mahout scale with the file size?
Figure 2. Total convergence times for: 5 Multi node cluster and Single
It can be observed on table 1 that a 5 multi node has node. File size: 1.1 GB.
worse performance than the single one for small datasets.
With a 66 MB file (6%) the distribution process is not B – How does Mahout scale with the number of nodes?
compensated by the overhead of the initial step. Given that
the HDFS block size is 64 MB, only two tasks will be Mahout decreases the clustering time as the number of
executed. Since each machine is a dual core there is no gain nodes increases. For 20 nodes it has reduced to less than one
expected for a file smaller than 128 MB. The gain is already fifth. In figure 3, it can be noticed that the cluster time
present for a file slightly bigger than 128 MB (12% diminishes with the number of nodes in a non linear way.
scenario). The overhead is largely compensated as soon as
the dataset grows. The gain in performance reaches 351%
for a 1.1 GB file.

Data File kmeans kmeans Iterations Gain
(%) MN (sec) SN (sec) to converge (%)
6 43.9 41.3 72 -6%
12 45.3 52.8 78 16%
25 46.4 88.5 72 91%
50 48.6 149.1 71 207%
100 70.3 316.7 56 351%
Table 1. Times per iteration and gain with the file size. MN - 5 Multi node.
SN - Single node.

Figure 2 shows that when the data file is increased, the 5 Figure 3. Runtimes per iteration with the number of nodes. File size: 1.1
GB.
node time enlarges very slowly. This seems to indicate that
the initial step has a considerable footprint. For this specific
dataset and set of fixed initial centroids, the K-means One can verify on table 2 that the gain is almost 100%
needed fewer interactions to converge as the file became between 1 and 2 nodes. However, it increases only 444%
bigger. Since the k is the same this is an expected behavior. between 1 and 20 nodes. This happens due to the size of the
Given that there are more samples per cluster, the averages dataset, which seems to be too small for a 20 nodes. The file
of centroids are more stable and the algorithm converges is divided into 18 blocks. It was noticed 2 idle machines in
with less interactions. However, each iteration is more the 20 nodes cluster. For this dataset a possible improvement
computer demanding. On the other side, more iterations could be forcing a higher number of tasks.
means more communication between the mappers and the It is interesting to notice that when the number of nodes
reducers. On a cluster this implies more network usage, increases, the CPU usage diminishes and the network usage
which can signify a bottleneck. enlarges. This can be clearly seen comparing figure 4, 5 and
6, 7. Those graphics were obtained from Amazon
Cloudwatch and correspond to the monitoring of a slave
node. They compare the situation between a 5 and a 10
nodes cluster executing K-means with the full dataset. The
CPU idle time on the 10 nodes scenario is noticeable. It
indicates that the instance was waiting for the results from
the other nodes. The network was continually used around

517

84 Mbps in this scenario while it was interrupted on the 5
nodes. The existence of extremely high peaks was present
when the cluster was not executing a job. As so, they should
be related to Amazon and not to the experiment itself. Both

Bytes
graphs show how intensive was the network usage. They
also point towards a saturation of the network usage.

Nodes K-means Gain
(min) (%)
1 296
2 149 98%
5 80 271% Figure 7. Network usage per node. 10 nodes.
File size: 1.1 GB.
10 66 351%
20 54 444%
Table 2. Converging times varying the number of nodes. C – How is Mahout influenced by the distance measure?
File size: 1.1 GB. Gain comparing with 1 node.
The Euclidean Squared distance metric uses the same
equation as the Euclidean distance metric, but does not take
the square root. As a result, clustering with the Euclidean
Squared distance metric should be considerably faster than
Utilization %

clustering with the regular Euclidean distance. However, the
experiment shows just a small improvement.

Distance K-means Iteractions
measure (min)
Euclidean 80 56
Squared Euclidean 79 56
Figure 4. CPU utilization per slave node. xx axis represents the time when Table 3. Converging times using different measures. Cluster size 5.
the experiment occured.5 nodes. File size: 1.1 GB. File size: 1.1 GB.

As K-means in table 3 show the squared Euclidean
reduces only 1 minute to 80. The difference between the
distances would reduce each map and reducer times. This
Utilization %

evidence together with others presented in this work
suggests that the first phase and the communication between
the machines can be relevant for the performance of
Mahout. It also implies that the reducer may have a
significant impact on each iteration time.

Figure 5. CPU utilization per slave node. 10 nodes. V. FUTURE WORK
File size: 1.1 GB.

In the future the following influences on Mahout K-
means will be studied:
• bigger datasets;
• increase in the number of tasks;
• increase in the number of reducers;
Bytes

• one machine with n cores vs n machines with 1
core.

VI. CONCLUSIONS
Figure 6. Network usage per node. 5 nodes.
File size: 1.1 GB. In this paper we tested how Mahout K-means scale. For
that purpose we have conducted several experiments

518

varying the dataset size and the number of nodes. The [7] S. B. Kotsiantis , P.E.P.: ‘Recent Advances in Clustering: A Brief
Survey’, WSEAS Transactions on Information Science and
influence of two different distance measures was also tested. Applications, 2004, 1, pp. 73 - 81
We concluded that Mahout can be a promising tool for K- [8] Abbas, O.A.: ‘Comparisons Between Data Clustering Algorithms’,
means clustering. It allows the clustering of large datasets, is The International Arab Journal of Information Technology, Vol 5.,
scalable and it produces significant gains in performance. No. 3, July 2008, 2008, 5, (3), pp. 320-325
Using Mahout with small files is not always the best [9] Singla, B., Yadav, K., and Singh, J.: ‘Comparison and analysis of
option. There is an overhead due to replication and clustering techniques’, in Editor (Ed.)^(Eds.): ‘Book Comparison and
analysis of clustering techniques’ (2008, edn.), pp. 1-3
distribution of the data blocks. The overhead is largely
[10] YING ZHAO, G.K.: ‘Hierarchical Clustering Algorithms for
compensated as soon as the dataset grows. Document Datasets’, Data Mining and Knowledge Discovery, 2005,
Increasing the number of nodes reduces the execution 10, pp. 141-168
times. However, for small files it can lead to an [11] Frigui, H., and Krishnapuram, R.: ‘Clustering by competitive
underutilization of each machine’s resources. agglomeration’, Pattern Recognition, 1997, 30, (7), pp. 1109-1119
The converging times do not depend largely on the [12] Pang-Ning Tan, M.S.a.V.K.: ‘Introduction to Data Mining ’
distance measure. (Addison-Wesley Longman Publishing Co., Inc., 2005, 1st edn. 2005)
As expected the quality of Mahout K-means clustering [13] C. IMMACULATE MARY, D.S.V.K.R.: ‘Refinement of clusters
from k-means with ant colony optimization’, Journal of Theoretical
was consistent between the several experiments. and Applied Information Technology 2009, 9. No. 2
[14] Dean, J., and Ghemawat, S.: ‘MapReduce: simplified data processing
on large clusters’, Commun. ACM, 2008, 51, (1), pp. 107-113
REFERENCES [15] White, T.: ‘Hadoop: The Definitive Guide’ (O'Reilly Media, 2009)
[1] Fredrik Farnstrom, J.L.a.C.E.: ‘Scalability for Clustering Algorithms [16] Sean Owen, R.A., Ted Dunning and Ellen Friedman: ‘Mahout in
Revisited’, SIGKDD Explorations, 2000, 2, pp. 51-57 Action’ (Manning Publications, 2010. 2010)
[2] Jiawei Han, M.K., Jian Pei: ‘Data Mining: Concepts and Techniques’ [17] Raskutti, B., and Leckie, C.: ‘An Evaluation of Criteria for Measuring
(Elsevier, 2006, 2nd edn. 2006) the Quality of Clusters’. Proc. Proceedings of the Sixteenth
[3] Osei-Bryson, K.-M.: ‘Assessing Cluster Quality Using Multiple International Joint Conference on Artificial Intelligence1999 pp.
Measures - A Decision Tree Based Approach in The Next Wave in Pages
Computing’, The Next Wave in Computing, Optimization, and [18] Rayward-Smith, Q.H.N.a.V.J.: ‘Internal quality measures for
Decision Technologies, 2005 29, (VI), pp. 371-384 clustering in metric space’, Int. J. Business Intelligence and Data
[4] Kun Niu, C.H., Shubo Zhang, and Junliang Chen: ‘ODDC: Outlier Mining 2008, 3, (1)
Detection Using Distance Distribution Clustering’. Proc. PAKDD [19] Shang, W., Adams, B., and Hassan, A.E.: ‘An experience report on
2007 Workshops, 2007 2007 pp. Pages scaling tools for mining software repositories using MapReduce’.
[5] Austin, V.J.H.a.J.: ‘A Survey of Outlier Detection Methodologies’, Proc. Proceedings of the IEEE/ACM international conference on
Artificial Intelligence Review, 2004, 22, pp. 85-126 Automated software engineering, Antwerp, Belgium pp. Pages
[6] A.K. JAIN, M.N.M.a.P.J.F.: ‘Data Clustering: A Review’, ACM [20] Cheng T. Chu, S.K.K., Yi A. Lin, Yuanyuan Yu, Gary R. Bradski,
Computing Surveys, 1999, 31, No. 3, September Andrew Y. Ng, Kunle Olukotun: ‘Map-Reduce for Machine Learning
on Multicore’, NIPS, 2006, pp. 281-288

519

K means clustering in the cloud - a mahout test

More Related Content

What's hot (18)

Similar to K means clustering in the cloud - a mahout test (20)

More from João Gabriel Lima (20)

Recently uploaded (20)

K means clustering in the cloud - a mahout test