Exploiting dynamic resource allocation for

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS 1

Exploiting Dynamic Resource Allocation for
Efficient Parallel Data Processing in the Cloud
Daniel Warneke and Odej Kao

Abstract—In recent years ad-hoc parallel data processing has emerged to be one of the killer applications for Infrastructure-as-a-
Service (IaaS) clouds. Major Cloud computing companies have started to integrate frameworks for parallel data processing in their
product portfolio, making it easy for customers to access these services and to deploy their programs. However, the processing
frameworks which are currently used have been designed for static, homogeneous cluster setups and disregard the particular nature
of a cloud. Consequently, the allocated compute resources may be inadequate for big parts of the submitted job and unnecessarily
increase processing time and cost. In this paper we discuss the opportunities and challenges for efficient parallel data processing
in clouds and present our research project Nephele. Nephele is the first data processing framework to explicitly exploit the dynamic
resource allocation offered by today’s IaaS clouds for both, task scheduling and execution. Particular tasks of a processing job can be
assigned to different types of virtual machines which are automatically instantiated and terminated during the job execution. Based on
this new framework, we perform extended evaluations of MapReduce-inspired processing jobs on an IaaS cloud system and compare
the results to the popular data processing framework Hadoop.

Index Terms—Many-Task Computing, High-Throughput Computing, Loosely Coupled Applications, Cloud Computing

!

1 I NTRODUCTION For companies that only have to process large amounts
Today a growing number of companies have to process of data occasionally running their own data center is
huge amounts of data in a cost-efficient manner. Classic obviously not an option. Instead, Cloud computing has
representatives for these companies are operators of emerged as a promising approach to rent a large IT in-
Internet search engines, like Google, Yahoo, or Microsoft. frastructure on a short-term pay-per-usage basis. Opera-
The vast amount of data they have to deal with ev- tors of so-called Infrastructure-as-a-Service (IaaS) clouds,
ery day has made traditional database solutions pro- like Amazon EC2 [1], let their customers allocate, access,
hibitively expensive [5]. Instead, these companies have and control a set of virtual machines (VMs) which run
popularized an architectural paradigm based on a large inside their data centers and only charge them for the
number of commodity servers. Problems like processing period of time the machines are allocated. The VMs are
crawled documents or regenerating a web index are split typically offered in different types, each type with its
into several independent subtasks, distributed among own characteristics (number of CPU cores, amount of
the available nodes, and computed in parallel. main memory, etc.) and cost.
In order to simplify the development of distributed Since the VM abstraction of IaaS clouds fits the ar-
applications on top of such architectures, many of these chitectural paradigm assumed by the data processing
companies have also built customized data processing frameworks described above, projects like Hadoop [25],
frameworks. Examples are Google’s MapReduce [9], Mi- a popular open source implementation of Google’s
crosoft’s Dryad [14], or Yahoo!’s Map-Reduce-Merge [6]. MapReduce framework, already have begun to promote
They can be classified by terms like high throughput using their frameworks in the cloud [29]. Only recently,
computing (HTC) or many-task computing (MTC), de- Amazon has integrated Hadoop as one of its core in-
pending on the amount of data and the number of frastructure services [2]. However, instead of embracing
tasks involved in the computation [20]. Although these its dynamic resource allocation, current data processing
systems differ in design, their programming models frameworks rather expect the cloud to imitate the static
share similar objectives, namely hiding the hassle of nature of the cluster environments they were originally
parallel programming, fault tolerance, and execution op- designed for. E.g., at the moment the types and number
timizations from the developer. Developers can typically of VMs allocated at the beginning of a compute job
continue to write sequential programs. The processing cannot be changed in the course of processing, although
framework then takes care of distributing the program the tasks the job consists of might have completely
among the available nodes and executes each instance different demands on the environment. As a result,
of the program on the appropriate fragment of data. rented resources may be inadequate for big parts of the
processing job, which may lower the overall processing
• The authors are with the Berlin University of Technology, Einsteinufer 17, performance and increase the cost.
10587 Berlin, Germany. In this paper we want to discuss the particular chal-
E-mail: daniel.warneke@tu-berlin.de, odej.kao@tu-berlin.de
lenges and opportunities for efficient parallel data pro-
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


cessing in clouds and present Nephele, a new processing dynamically and just for the time they are required in
framework explicitly designed for cloud environments. the processing workflow. E.g., a framework exploiting
Most notably, Nephele is the first data processing frame- the possibilities of a cloud could start with a single VM
work to include the possibility of dynamically allo- which analyzes an incoming job and then advises the
cating/deallocating different compute resources from a cloud to directly start the required VMs according to the
cloud in its scheduling and during job execution. job’s processing phases. After each phase, the machines
This paper is an extended version of [27]. It includes could be released and no longer contribute to the overall
further details on scheduling strategies and extended cost for the processing job.
experimental results. The paper is structured as fol- Facilitating such use cases imposes some requirements
lows: Section 2 starts with analyzing the above men- on the design of a processing framework and the way its
tioned opportunities and challenges and derives some jobs are described. First, the scheduler of such a frame-
important design principles for our new framework. In work must become aware of the cloud environment a job
Section 3 we present Nephele’s basic architecture and should be executed in. It must know about the different
outline how jobs can be described and executed in the types of available VMs as well as their cost and be able to
cloud. Section 4 provides some first figures on Nephele’s allocate or destroy them on behalf of the cloud customer.
performance and the impact of the optimizations we Second, the paradigm used to describe jobs must be
propose. Finally, our work is concluded by related work powerful enough to express dependencies between the
(Section 5) and ideas for future work (Section 6). different tasks the jobs consists of. The system must
be aware of which task’s output is required as another
2 C HALLENGES AND O PPORTUNITIES task’s input. Otherwise the scheduler of the processing
framework cannot decide at what point in time a par-
Current data processing frameworks like Google’s
ticular VM is no longer needed and deallocate it. The
MapReduce or Microsoft’s Dryad engine have been de-
MapReduce pattern is a good example of an unsuitable
signed for cluster environments. This is reflected in a
paradigm here: Although at the end of a job only few
number of assumptions they make which are not nec-
reducer tasks may still be running, it is not possible to
essarily valid in cloud environments. In this section we
shut down the idle VMs, since it is unclear if they contain
discuss how abandoning these assumptions raises new
intermediate results which are still required.
opportunities but also challenges for efficient parallel
Finally, the scheduler of such a processing framework
data processing in clouds.
must be able to determine which task of a job should
be executed on which type of VM and, possibly, how
2.1 Opportunities many of those. This information could be either provided
Today’s processing frameworks typically assume the re- externally, e.g. as an annotation to the job description,
sources they manage consist of a static set of homogeneous or deduced internally, e.g. from collected statistics, sim-
compute nodes. Although designed to deal with individ- ilarly to the way database systems try to optimize their
ual nodes failures, they consider the number of available execution schedule over time [24].
machines to be constant, especially when scheduling
the processing job’s execution. While IaaS clouds can 2.2 Challenges
certainly be used to create such cluster-like setups, much The cloud’s virtualized nature helps to enable promising
of their flexibility remains unused. new use cases for efficient parallel data processing. How-
One of an IaaS cloud’s key features is the provisioning ever, it also imposes new challenges compared to classic
of compute resources on demand. New VMs can be cluster setups. The major challenge we see is the cloud’s
allocated at any time through a well-defined interface opaqueness with prospect to exploiting data locality:
and become available in a matter of seconds. Machines In a cluster the compute nodes are typically intercon-
which are no longer used can be terminated instantly nected through a physical high-performance network.
and the cloud customer will be charged for them no The topology of the network, i.e. the way the compute
more. Moreover, cloud operators like Amazon let their nodes are physically wired to each other, is usually well-
customers rent VMs of different types, i.e. with different known and, what is more important, does not change
computational power, different sizes of main memory, over time. Current data processing frameworks offer to
and storage. Hence, the compute resources available in leverage this knowledge about the network hierarchy
a cloud are highly dynamic and possibly heterogeneous. and attempt to schedule tasks on compute nodes so that
With respect to parallel data processing, this flexibility data sent from one node to the other has to traverse as
leads to a variety of new possibilities, particularly for few network switches as possible [9]. That way network
scheduling data processing jobs. The question a sched- bottlenecks can be avoided and the overall throughput
uler has to answer is no longer “Given a set of compute of the cluster can be improved.
resources, how to distribute the particular tasks of a job In a cloud this topology information is typically not
among them?”, but rather “Given a job, what compute exposed to the customer [29]. Since the nodes involved
resources match the tasks the job consists of best?”. in processing a data intensive job often have to trans-
This new paradigm allows allocating compute resources fer tremendous amounts of data through the network,
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


this drawback is particularly severe; parts of the net- Manager (JM). The Job Manager receives the client’s
work may become congested while others are essen- jobs, is responsible for scheduling them, and coordinates
tially unutilized. Although there has been research on their execution. It is capable of communicating with
inferring likely network topologies solely from end- the interface the cloud operator provides to control the
to-end measurements (e.g. [7]), it is unclear if these instantiation of VMs. We call this interface the Cloud
techniques are applicable to IaaS clouds. For security Controller. By means of the Cloud Controller the Job
reasons clouds often incorporate network virtualization Manager can allocate or deallocate VMs according to the
techniques (e.g. [8]) which can hamper the inference pro- current job execution phase. We will comply with com-
cess, in particular when based on latency measurements. mon Cloud computing terminology and refer to these
Even if it was possible to determine the underlying VMs as instances for the remainder of this paper. The
network hierarchy in a cloud and use it for topology- term instance type will be used to differentiate between
aware scheduling, the obtained information would not VMs with different hardware characteristics. E.g., the
necessarily remain valid for the entire processing time. instance type “m1.small” could denote VMs with one
VMs may be migrated for administrative purposes be- CPU core, one GB of RAM, and a 128 GB disk while the
tween different locations inside the data center without instance type “c1.xlarge” could refer to machines with 8
any notification, rendering any previous knowledge of CPU cores, 18 GB RAM, and a 512 GB disk.
the relevant network infrastructure obsolete. The actual execution of tasks which a Nephele job
As a result, the only way to ensure locality between consists of is carried out by a set of instances. Each
tasks of a processing job is currently to execute these instance runs a so-called Task Manager (TM). A Task
tasks on the same VM in the cloud. This may involve Manager receives one or more tasks from the Job Man-
allocating fewer, but more powerful VMs with multiple ager at a time, executes them, and after that informs the
CPU cores. E.g., consider an aggregation task receiving Job Manager about their completion or possible errors.
data from seven generator tasks. Data locality can be Unless a job is submitted to the Job Manager, we expect
ensured by scheduling these tasks to run on a VM with the set of instances (and hence the set of Task Managers)
eight cores instead of eight distinct single-core machines. to be empty. Upon job reception the Job Manager then
However, currently no data processing framework in- decides, depending on the job’s particular tasks, how
cludes such strategies in its scheduling algorithms. many and what type of instances the job should be
executed on, and when the respective instances must
3 D ESIGN be allocated/deallocated to ensure a continuous but
cost-efficient processing. Our current strategies for these
Based on the challenges and opportunities outlined
decisions are highlighted at the end of this section.
in the previous section we have designed Nephele, a
The newly allocated instances boot up with a previ-
new data processing framework for cloud environments.
ously compiled VM image. The image is configured to
Nephele takes up many ideas of previous processing
automatically start a Task Manager and register it with
frameworks but refines them to better match the dy-
the Job Manager. Once all the necessary Task Managers
namic and opaque nature of a cloud.
have successfully contacted the Job Manager, it triggers
the execution of the scheduled job.
3.1 Architecture Initially, the VM images used to boot up the Task
Nephele’s architecture follows a classic master-worker Managers are blank and do not contain any of the data
pattern as illustrated in Fig. 1. the Nephele job is supposed to operate on. As a result,
we expect the cloud to offer persistent storage (like
Client e.g. Amazon S3 [3]). This persistent storage is supposed
to store the job’s input data and eventually receive
Public Network (Internet) its output data. It must be accessible for both the Job
Cloud Manager as well as for the set of Task Managers, even
Cloud Controller

JM if they are connected by a private or virtual network.
Persistent Storage

Private / Virtualized Network
3.2 Job description
Similar to Microsoft’s Dryad [14], jobs in Nephele are
TM TM TM TM expressed as a directed acyclic graph (DAG). Each vertex
in the graph represents a task of the overall processing
job, the graph’s edges define the communication flow
Fig. 1. Structural overview of Nephele running in an between these tasks. We also decided to use DAGs to
Infrastructure-as-a-Service (IaaS) cloud describe processing jobs for two major reasons:
The first reason is that DAGs allow tasks to have
Before submitting a Nephele compute job, a user must multiple input and multiple output edges. This tremen-
start a VM in the cloud which runs the so called Job dously simplifies the implementation of classic data
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


combining functions like, e.g., join operations [6]. Sec- plicity: Users should be able to describe tasks and their
ond and more importantly, though, the DAG’s edges relationships on an abstract level. Therefore, the Job
explicitly model the communication paths of the process- Graph does not explicitly model task parallelization and
ing job. As long as the particular tasks only exchange the mapping of tasks to instances. However, users who
data through these designated communication edges, wish to influence these aspects can provide annotations
Nephele can always keep track of what instance might to their job description. These annotations include:
still require data from what other instances and which • Number of subtasks: A developer can declare his
instance can potentially be shut down and deallocated. task to be suitable for parallelization. Users that
Defining a Nephele job comprises three mandatory include such tasks in their Job Graph can specify
steps: First, the user must write the program code for how many parallel subtasks Nephele should split
each task of his processing job or select it from an exter- the respective task into at runtime. Subtasks execute
nal library. Second, the task program must be assigned the same task code, however, they typically process
to a vertex. Finally, the vertices must be connected by different fragments of the data.
edges to define the communication paths of the job. • Number of subtasks per instance: By default each
Tasks are expected to contain sequential code and pro- subtask is assigned to a separate instance. In case
cess so-called records, the primary data unit in Nephele. several subtasks are supposed to share the same
Programmers can define arbitrary types of records. From instance, the user can provide a corresponding an-
a programmer’s perspective records enter and leave the notation with the respective task.
task program through input or output gates. Those input • Sharing instances between tasks: Subtasks of dif-
and output gates can be considered endpoints of the ferent tasks are usually assigned to different (sets of)
DAG’s edges which are defined in the following step. instances unless prevented by another scheduling
Regular tasks (i.e. tasks which are later assigned to inner restriction. If a set of instances should be shared
vertices of the DAG) must have at least one or more between different tasks the user can attach a cor-
input and output gates. Contrary to that, tasks which responding annotation to the Job Graph.
either represent the source or the sink of the data flow • Channel types: For each edge connecting two ver-
must not have input or output gates, respectively. tices the user can determine a channel type. Before
After having specified the code for the particular tasks executing a job, Nephele requires all edges of the
of the job, the user must define the DAG to connect these original Job Graph to be replaced by at least one
tasks. We call this DAG the Job Graph. The Job Graph channel of a specific type. The channel type dictates
maps each task to a vertex and determines the commu- how records are transported from one subtask to
nication paths between them. The number of a vertex’s another at runtime. Currently, Nephele supports
incoming and outgoing edges must thereby comply with network, file, and in-memory channels. The choice
the number of input and output gates defined inside the of the channel type can have several implications on
tasks. In addition to the task to execute, input and output the entire job schedule. A more detailled disussion
vertices (i.e. vertices with either no incoming or outgoing on this is provided in the next subsection.
edge) can be associated with a URL pointing to external • Instance type: A subtask can be executed on dif-
storage facilities to read or write input or output data, ferent instance types which may be more or less
respectively. Figure 2 illustrates the simplest possible Job suitable for the considered program. Therefore we
Graph. It only consists of one input, one task, and one have developed special annotations task developers
output vertex. can use to characterize the hardware requirements
of their code. However, a user who simply utilizes
these annotated tasks can also overwrite the devel-
Output 1
oper’s suggestion and explicitly specify the instance
Task: LineWriterTask.program
Output: s3://user:key@storage/outp type for a task in the Job Graph.
If the user omits to augment the Job Graph with
these specifications, Nephele’s scheduler applies default
Task 1 strategies which are discussed later on in this section.
Task: MyTask.program Once the Job Graph is specified, the user submits it to
the Job Manager, together with the credentials he has
obtained from his cloud operator. The credentials are
Input 1 required since the Job Manager must allocate/deallocate
Task: LineReaderTask.program
instances during the job execution on behalf of the user.
Input: s3://user:key@storage/input

3.3 Job Scheduling and Execution
Fig. 2. An example of a Job Graph in Nephele After having received a valid Job Graph from the user,
Nephele’s Job Manager transforms it into a so-called Ex-
One major design goal of Job Graphs has been sim- ecution Graph. An Execution Graph is Nephele’s primary
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


data structure for scheduling and monitoring the execu- cloud, Nephele allows to allocate/deallocate instances in
tion of a Nephele job. Unlike the abstract Job Graph, the the course of the processing job, when some subtasks
Execution Graph contains all the concrete information have already been completed or are already running.
required to schedule and execute the received job on However, this just-in-time allocation can also cause prob-
the cloud. It explicitly models task parallelization and lems, since there is the risk that the requested instance
the mapping of tasks to instances. Depending on the types are temporarily not available in the cloud. To
level of annotations the user has provided with his Job cope with this problem, Nephele separates the Execution
Graph, Nephele may have different degrees of freedom Graph into one or more so-called Execution Stages. An
in constructing the Execution Graph. Figure 3 shows one Execution Stage must contain at least one Group Vertex.
possible Execution Graph constructed from the previ- Its processing can only start when all the subtasks in-
ously depicted Job Graph (Figure 2). Task 1 is e.g. split cluded in the preceding stages have been successfully
into two parallel subtasks which are both connected to processed. Based on this Nephele’s scheduler ensures
the task Output 1 via file channels and are all scheduled the following three properties for the entire job execu-
to run on the same instance. The exact structure of the tion: First, when the processing of a stage begins, all
Execution Graph is explained in the following: instances required within the stage are allocated. Second,
all subtasks included in this stage are set up (i.e. sent
Stage 1 to the corresponding Task Managers along with their
Execution
Output 1 (1) Stage required libraries) and ready to receive records. Third,
ID: i-40A608A3 Group
before the processing of a new stage, all intermediate
Type: m1.large Vertex results of its preceding stages are stored in a persistent
manner. Hence, Execution Stages can be compared to
Execution checkpoints. In case a sufficient number of resources
Stage 0 Instance
cannot be allocated for the next stage, they allow a
Execution running job to be interrupted and later on restored when
Vertex
enough spare resources have become available.
Task 1 (2) The concrete level of the Execution Graph refines the
job schedule to include subtasks and their communica-
ID: i-59BC0013 tion channels. In Nephele, every task is transformed into
Type: m1.small either exactly one, or, if the task is suitable for parallel
Input 1 (1)
execution, at least one subtask. For a task to complete
successfully, each of its subtasks must be successfully
File channel Network channel
processed by a Task Manager. Subtasks are represented
by so-called Execution Vertices in the Execution Graph.
Fig. 3. An Execution Graph created from the original Job They can be considered the most fine-grained executable
Graph job unit. To simplify management, each Execution Vertex
is always controlled by its corresponding Group Vertex.
In contrast to the Job Graph, an Execution Graph is Nephele allows each task to be executed on its own
no longer a pure DAG. Instead, its structure resembles instance type, so the characteristics of the requested
a graph with two different levels of details, an abstract VMs can be adapted to the demands of the current
and a concrete level. While the abstract graph describes processing phase. To reflect this relation in the Execution
the job execution on a task level (without parallelization) Graph, each subtask must be mapped to a so-called
and the scheduling of instance allocation/deallocation, Execution Instance. An Execution Instance is defined by
the concrete, more fine-grained graph defines the map- an ID and an instance type representing the hardware
ping of subtasks to instances and the communication characteristics of the corresponding VM. It is a schedul-
channels between them. ing stub that determines which subtasks have to run
On the abstract level, the Execution Graph equals on what instance (type). We expect a list of available
the user’s Job Graph. For every vertex of the original instance types together with their cost per time unit to be
Job Graph there exists a so-called Group Vertex in the accessible for Nephele’s scheduler and instance types to
Execution Graph. As a result, Group Vertices also rep- be referable by simple identifier strings like “m1.small”.
resent distinct tasks of the overall job, however, they Before processing a new Execution Stage, the sched-
cannot be seen as executable units. They are used as a uler collects all Execution Instances from that stage and
management abstraction to control the set of subtasks the tries to replace them with matching cloud instances. If
respective task program is split into. The edges between all required instances could be allocated the subtasks are
Group Vertices are only modeled implicitly as they do distributed among them and set up for execution.
not represent any physical communication paths during On the concrete level, the Execution Graph inherits the
the job processing. For the sake of presentation, they are edges from the abstract level, i.e. edges between Group
also omitted in Figure 3. Vertices are translated into edges between Execution Ver-
In order to ensure cost-efficient execution in an IaaS tices. In case of task parallelization, when a Group Vertex
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


contains more than one Execution Vertex, the developer a simple default strategy: Each vertex of the Job Graph is
of the consuming task can implement an interface which transformed into one Execution Vertex. The default chan-
determines how to connect the two different groups nel types are network channels. Each Execution Vertex is
of subtasks. The actual number of channels that are by default assigned to its own Execution Instance unless
connected to a subtask at runtime is hidden behind the the user’s annotations or other scheduling restrictions
task’s respective input and output gates. However, the (e.g. the usage of in-memory channels) prohibit it. The
user code can determine the number if necessary. default instance type to be used is the one with the
Nephele requires all edges of an Execution Graph to lowest price per time unit available in the IaaS cloud.
be replaced by a channel before processing can begin. One fundamental idea to refine the scheduling strat-
The type of the channel determines how records are egy for recurring jobs is to use feedback data. We de-
transported from one subtask to the other. Currently, veloped a profiling subsystem for Nephele which can
Nephele features three different types of channels, which continously monitor running tasks and the underlying
all put different constrains on the Execution Graph. instances. Based on the Java Management Extensions
• Network channels: A network channel lets two sub- (JMX) the profiling subsystem is, among other things,
tasks exchange data via a TCP connection. Network capable of breaking down what percentage of its pro-
channels allow pipelined processing, so the records cessing time a task thread actually spends processing
emitted by the producing subtask are immediately user code and what percentage of time it has to wait for
transported to the consuming subtask. As a result, data. With the collected data Nephele is able to detect
two subtasks connected via a network channel may both computational as well as I/O bottlenecks. While
be executed on different instances. However, since computational bottlenecks suggest a higher degree of
they must be executed at the same time, they are parallelization for the affected tasks, I/O bottlenecks
required to run in the same Execution Stage. provide hints to switch to faster channel types (like in-
• In-Memory channels: Similar to a network chan- memory channels) and reconsider the instance assign-
nel, an in-memory channel also enables pipelined ment. Since Nephele calculates a cryptographic signature
processing. However, instead of using a TCP con- for each task, recurring tasks can be identified and the
nection, the respective subtasks exchange data using previously recorded feedback data can be exploited.
the instance’s main memory. An in-memory channel At the moment we only use the profiling data to detect
typically represents the fastest way to transport these bottlenecks and help the user to choose reasonable
records in Nephele, however, it also implies most annotations for his job. Figure 4 illustrates the graphical
scheduling restrictions: The two connected subtasks job viewer we have devised for that purpose. It provides
must be scheduled to run on the same instance and immediate visual feedback about the current utilization
run in the same Execution Stage. of all tasks and cloud instances involved in the compu-
• File channels: A file channel allows two subtasks tation. A user can utilize this visual feedback to improve
to exchange records via the local file system. The his job annotations for upcoming job executions. In more
records of the producing task are first entirely writ- advanced versions of Nephele we envision the system
ten to an intermediate file and afterwards read to automatically adapt to detected bottlenecks, either
into the consuming subtask. Nephele requires two between consecutive executions of the same job or even
such subtasks to be assigned to the same instance. during job execution at runtime.
Moreover, the consuming Group Vertex must be While the allocation time of cloud instances is deter-
scheduled to run in a higher Execution Stage than mined by the start times of the assigned subtasks, there
the producing Group Vertex. In general, Nephele are different possible strategies for instance deallocation.
only allows subtasks to exchange records across In order to reflect the fact that most cloud providers
different stages via file channels because they are charge their customers for instance usage by the hour, we
the only channel types which store the intermediate integrated the possibility to reuse instances. Nephele can
records in a persistent manner. keep track of the instances’ allocation times. An instance
of a particular type which has become obsolete in the
current Execution Stage is not immediately deallocated if
3.4 Parallelization and Scheduling Strategies an instance of the same type is required in an upcoming
As mentioned before, constructing an Execution Graph Execution Stage. Instead, Nephele keeps the instance
from a user’s submitted Job Graph may leave different allocated until the end of its current lease period. If the
degrees of freedom to Nephele. Using this freedom to next Execution Stage has begun before the end of that
construct the most efficient Execution Graph (in terms of period, it is reassigned to an Execution Vertex of that
processing time or monetary cost) is currently a major stage, otherwise it deallocated early enough not to cause
focus of our research. Discussing this subject in detail any additional cost.
would go beyond the scope of this paper. However, we Besides the use of feedback data we recently com-
want to outline our basic approaches in this subsection: plemented our efforts to provide reasonable job an-
Unless the user provides any job annotation which notations automatically by a higher-level programming
contains more specific instructions we currently pursue model layered on top of Nephele. Rather than describing
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


Fig. 4. Nephele’s job viewer provides graphical feedback on instance utilization and detected bottlenecks

jobs as arbitrary DAGs, this higher-level programming numbers. The second task subsequently is to calculate
model called PACTs [4] is centered around the con- the average of these k smallest numbers. The job is
catenation of second-order functions, e.g. like the map a classic representative for a variety of data analysis
and reduce function from the well-known MapReduce jobs whose particular tasks vary in their complexity
programming model. Developers can write custom first- and hardware demands. While the first task has to sort
order functions and attach them to the desired second- the entire data set and therefore can take advantage of
order functions. The PACTs programming model is se- large amounts of main memory and parallel execution,
mantically richer than Nephele’s own programming ab- the second aggregation task requires almost no main
straction. E.g. it considers aspects like the input/output memory and, at least eventually, cannot be parallelized.
cardinalities of the first order functions which is helpful
to deduce reasonable degrees of parallelization. More We implemented the described sort/aggregate task
details can be found in [4]. for three different experiments. For the first experiment,
we implemented the task as a sequence of MapReduce
programs and executed it using Hadoop on a fixed set
4 E VALUATION of instances. For the second experiment, we reused the
In this section we want to present first performance same MapReduce programs as in the first experiment
results of Nephele and compare them to the data pro- but devised a special MapReduce wrapper to make
cessing framework Hadoop. We have chosen Hadoop these programs run on top of Nephele. The goal of
as our competitor, because it is an open source soft- this experiment was to illustrate the benefits of dynamic
ware and currently enjoys high popularity in the data resource allocation/deallocation while still maintaining
processing community. We are aware that Hadoop has the MapReduce processing pattern. Finally, as the third
been designed to run on a very large number of nodes experiment, we discarded the MapReduce pattern and
(i.e. several thousand nodes). However, according to our implemented the task based on a DAG to also highlight
observations, the software is typically used with signif- the advantages of using heterogeneous instances.
icantly fewer instances in current IaaS clouds. In fact,
Amazon itself limits the number of available instances For all three experiments, we chose the data set size to
for their MapReduce service to 20 unless the respective be 100 GB. Each integer number had the size of 100 bytes.
customer passes an extended registration process [2]. As a result, the data set contained about 109 distinct
The challenge for both frameworks consists of two integer numbers. The cut-off variable k has been set to
abstract tasks: Given a set of random integer numbers, 2 · 108 , so the smallest 20 % of all numbers had to be
the first task is to determine the k smallest of those determined and aggregated.
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


4.1 General hardware setup the selected files and emitted the first 2 · 108 numbers
All three experiments were conducted on our local IaaS to the reducer. In order to enable parallelization in the
cloud of commodity servers. Each server is equipped reduce phase, we chose the intermediate keys for the
with two Intel Xeon 2.66 GHz CPUs (8 CPU cores) reducer randomly from a predefined set of keys. These
and a total main memory of 32 GB. All servers are keys ensured that the emitted numbers were distributed
connected through regular 1 GBit/s Ethernet links. The evenly among the n reducers in the system. Each reducer
8
host operating system was Gentoo Linux (kernel version then calculated the average of the received 2·10 integer
n
2.6.30) with KVM [15] (version 88-r1) using virtio [23] to numbers. The third MapReduce job finally read the n
provide virtual I/O access. intermediate average values and aggregated them to a
To manage the cloud and provision VMs on request of single overall average.
Nephele, we set up Eucalyptus [16]. Similar to Amazon Since Hadoop is not designed to deal with heteroge-
EC2, Eucalyptus offers a predefined set of instance types neous compute nodes, we allocated six instances of type
a user can choose from. During our experiments we used “c1.xlarge” for the experiment. All of these instances
two different instance types: The first instance type was were assigned to Hadoop throughout the entire duration
“m1.small” which corresponds to an instance with one of the experiment.
CPU core, one GB of RAM, and a 128 GB disk. The We configured Hadoop to perform best for the first,
second instance type, “c1.xlarge”, represents an instance computationally most expensive, MapReduce job: In ac-
with 8 CPU cores, 18 GB RAM, and a 512 GB disk. cordance to [18] we set the number of map tasks per job
Amazon EC2 defines comparable instance types and to 48 (one map task per CPU core) and the number of
offers them at a price of about 0.10 $, or 0.80 $ per hour reducers to 12. The memory heap of each map task as
(September 2009), respectively. well as the in-memory file system have been increased
The images used to boot up the instances contained to 1 GB and 512 MB, respectively, in order to avoid
Ubuntu Linux (kernel version 2.6.28) with no addi- unnecessarily spilling transient data to disk.
tional software but a Java runtime environment (version
1.6.0 13), which is required by Nephele’s Task Manager. 4.3 Experiment 2: MapReduce and Nephele
The 100 GB input data set of random integer numbers
has been generated according to the rules of the Jim Gray For the second experiment we reused the three MapRe-
sort benchmark [18]. In order to make the data accessible duce programs we had written for the previously de-
to Hadoop, we started an HDFS [25] data node on each scribed Hadoop experiment and executed them on top
of the allocated instances prior to the processing job of Nephele. In order to do so, we had to develop a
and distributed the data evenly among the nodes. Since set of wrapper classes providing limited interface com-
this initial setup procedure was necessary for all three patibility with Hadoop and sort/merge functionality.
experiments (Hadoop and Nephele), we have chosen to These wrapper classes allowed us to run the unmod-
ignore it in the following performance discussion. ified Hadoop MapReduce programs with Nephele. As
a result, the data flow was controlled by the executed
MapReduce programs while Nephele was able to govern
4.2 Experiment 1: MapReduce and Hadoop the instance allocation/deallocation and the assignment
In order to execute the described sort/aggregate task of tasks to instances during the experiment. We devised
with Hadoop we created three different MapReduce this experiment to highlight the effects of the dynamic
programs which were executed consecutively. resource allocation/deallocation while still maintaining
The first MapReduce job reads the entire input data comparability to Hadoop as well as possible.
set, sorts the contained integer numbers ascendingly, and Figure 5 illustrates the Execution Graph we instructed
writes them back to Hadoop’s HDFS file system. Since Nephele to create so that the communication paths
the MapReduce engine is internally designed to sort the match the MapReduce processing pattern. For brevity,
incoming data between the map and the reduce phase, we omit a discussion on the original Job Graph. Fol-
we did not have to provide custom map and reduce lowing our experiences with the Hadoop experiment,
functions here. Instead, we simply used the TeraSort we pursued the overall idea to also start with a ho-
code, which has recently been recognized for being well- mogeneous set of six “c1.xlarge” instances, but to re-
suited for these kinds of tasks [18]. The result of this duce the number of allocated instances in the course
first MapReduce job was a set of files containing sorted of the experiment according to the previously observed
integer numbers. Concatenating these files yielded the workload. For this reason, the sort operation should be
fully sorted sequence of 109 numbers. carried out on all six instances, while the first and second
The second and third MapReduce jobs operated on aggregation operation should only be assigned to two
the sorted data set and performed the data aggregation. instances and to one instance, respectively.
Thereby, the second MapReduce job selected the first The experiment’s Execution Graph consisted of three
output files from the preceding sort job which, just by Execution Stages. Each stage contained the tasks re-
their file size, had to contain the smallest 2 · 108 numbers quired by the corresponding MapReduce program. As
of the initial data set. The map function was fed with stages can only be crossed via file channels, all interme-
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


Stage 2
BigIntegerWriter (1)
ID: i-DB728EF1 In-memory channel
on 1 instance
Type: c1.xlarge
Network channel
AggregateReduce (1) File channel
on 1 instance

AggregateMap (4)
on 1 instance 4 subtasks

Stage 1
DummyTask (4)
on 1 instance 4 subtasks

AggregateReduce (4)
ID: i-DB728EF1 2 subtasks 2 subtasks ID: i-F3481EA2
on 2 instances
Type: c1.xlarge Type: c1.xlarge

AggregateMap (12)
ID: i-DB728EF1 ID: i-F3481EA2
Type: c1.xlarge ... 6 subtasks
... on 2 instances
... 6 subtasks ... Type: c1.xlarge

Stage 0

ID: i-DB728EF1 DummyTask (12) ID: i-F3481EA2
Type: c1.xlarge ... 6 subtasks ... on 2 instances ... 6 subtasks ... Type: c1.xlarge

TeraSortReduce (12)
Type: c1.xlarge
2 subtasks
... on 6 instances
... 2 subtasks
Type: c1.xlarge

TeraSortMap (48)
Type: c1.xlarge ... on 6 instances
... Type: c1.xlarge

... 8 subtasks ... BigIntegerReader (48) ... 8 subtasks ...
... on 6 instances
...
Fig. 5. The Execution Graph for Experiment 2 (MapReduce and Nephele)

diate data occurring between two succeeding MapRe- devised the following optimization: If two subtasks that
duce jobs was completely written to disk, like in the are connected by a network channel are scheduled to run
previous Hadoop experiment. on the same instance, their network channel is dynami-
The first stage (Stage 0) included four different tasks, cally converted into an in-memory channel. That way we
split into several different groups of subtasks and as- were able to avoid unnecessary data serialization and the
signed to different instances. The first task, BigInte- resulting processing overhead. For the given MapReduce
gerReader, processed the assigned input files and emitted communication pattern this optimization accounts for
each integer number as a separate record. The tasks approximately 20% less network channels.
TeraSortMap and TeraSortReduce encapsulated the Tera- The records emitted by the BigIntegerReader subtasks
Sort MapReduce code which had been executed by were received by the TeraSortMap subtasks. The TeraSort
Hadoop’s mapper and reducer threads before. In order partitioning function, located in the TeraSortMap task,
to meet the setup of the previous experiment, we split then determined which TeraSortReduce subtask was re-
the TeraSortMap and TeraSortReduce tasks into 48 and sponsible for the received record, depending on its value.
12 subtasks, respectively, and assigned them to six in- Before being sent to the respective reducer, the records
stances of type “c1.xlarge”. Furthermore, we instructed were collected in buffers of approximately 44 MB size
Nephele to construct network channels between each and were presorted in memory. Considering that each
pair of TeraSortMap and TeraSortReduce subtasks. For TeraSortMap subtask was connected to 12 TeraSortRe-
the sake of legibility, only few of the resulting channels duce subtasks, this added up to a buffer size of 574 MB,
are depicted in Figure 5. similar to the size of the in-memory file system we had
Although Nephele only maintains at most one phys- used for the Hadoop experiment previously.
ical TCP connection between two cloud instances, we Each TeraSortReducer subtask had an in-memory
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


buffer of about 512 MB size, too. The buffer was used tried to exploit Nephele’s ability to manage heteroge-
to mitigate hard drive access when storing the incom- neous compute resources.
ing sets of presorted records from the mappers. Like Figure 6 illustrates the Execution Graph we instructed
Hadoop, we started merging the first received presorted Nephele to create for this experiment. For brevity, we
record sets using separate background threads while the again leave out a discussion on the original Job Graph.
data transfer from the mapper tasks was still in progress. Similar to the previous experiment, we pursued the idea
To improve performance, the final merge step, resulting that several powerful but expensive instances are used to
in one fully sorted set of records, was directly streamed determine the 2·108 smallest integer numbers in parallel,
to the next task in the processing chain. while, after that, a single inexpensive instance is utilized
The task DummyTask, the forth task in the first Exe- for the final aggregation. The graph contained five dis-
cution Stage, simply emitted every record it received. tinct tasks, again split into different groups of subtasks.
It was used to direct the output of a preceding task to However, in contrast to the previous experiment, this
a particular subset of allocated instances. Following the one also involved instances of different types.
overall idea of this experiment, we used the DummyTask In order to feed the initial data from HDFS into
task in the first stage to transmit the sorted output of the Nephele, we reused the BigIntegerReader task. The
12 TeraSortReduce subtasks to the two instances Nephele records emitted by the BigIntegerReader subtasks were
would continue to work with in the second Execution received by the second task, BigIntegerSorter, which at-
Stage. For the DummyTask subtasks in the first stage tempted to buffer all incoming records into main mem-
(Stage 0) we shuffled the assignment of subtasks to in- ory. Once it had received all designated records, it
stances in a way that both remaining instances received performed an in-memory quick sort and subsequently
a fairly even fraction of the 2 · 108 smallest numbers. continued to emit the records in an order-preserving
Without the shuffle, the 2·108 would all be stored on only manner. Since the BigIntegerSorter task requires large
one of the remaining instances with high probability. amounts of main memory we split it into 146 sub-
In the second and third stage (Stage 1 and 2 in Fig- tasks and assigned these evenly to six instances of type
ure 5) we ran the two aggregation steps corresponding “c1.xlarge”. The preceding BigIntegerReader task was
to the second and third MapReduce program in the also split into 146 subtasks and set up to emit records
previous Hadoop experiment. AggregateMap and Aggre- via in-memory channels.
gateReduce encapsulated the respective Hadoop code. The third task, BigIntegerMerger, received records from
The first aggregation step was distributed across 12 multiple input channels. Once it has read a record from
AggregateMap and four AggregateReduce subtasks, as- all available input channels, it sorts the records locally
signed to the two remaining “c1.xlarge” instances. To and always emits the smallest number. The BigInte-
determine how many records each AggregateMap sub- gerMerger tasks occured three times in a row in the
task had to process, so that in total only the 2 · 108 Execution Graph. The first time it was split into six sub-
numbers would be emitted to the reducers, we had to tasks, one subtask assigned to each of the six “c1.xlarge”
develop a small utility program. This utility program instances. As described in Section 2, this is currently
consisted of two components. The first component ran the only way to ensure data locality between the sort
in the DummyTask subtasks of the preceding Stage 0. It and merge tasks. The second time the BigIntegerMerger
wrote the number of records each DummyTask subtask task occured in the Execution Graph, it was split into
had eventually emitted to a network file system share two subtasks. These two subtasks were assigned to two
which was accessible to every instance. The second of the previously used “c1.xlarge” instances. The third
component, integrated in the AggregateMap subtasks, occurrence of the task was assigned to new instance of
read those numbers and calculated what fraction of the the type “m1.small”.
sorted data was assigned to the respective mapper. In Since we abandoned the MapReduce processing pat-
the previous Hadoop experiment this auxiliary program tern, we were able to better exploit Nephele’s stream-
was unnecessary since Hadoop wrote the output of each ing pipelining characteristics in this experiment. Conse-
MapReduce job back to HDFS anyway. quently, each of the merge subtasks was configured to
After the first aggregation step, we again used the stop execution after having emitted 2 · 108 records. The
DummyTask task to transmit the intermediate results to stop command was propagated to all preceding subtasks
the last instance which executed the final aggregation in of the processing chain, which allowed the Execution
the third stage. The final aggregation was carried out by Stage to be interrupted as soon as the final merge subtask
four AggregateMap subtasks and one AggregateReduce had emitted the 2 · 108 smallest records.
subtask. Eventually, we used one subtask of BigInte- The fourth task, BigIntegerAggregater, read the incom-
gerWriter to write the final result record back to HDFS. ing records from its input channels and summed them
up. It was also assigned to the single “m1.small” in-
4.4 Experiment 3: DAG and Nephele stance. Since we no longer required the six “c1.xlarge”
In this third experiment we were no longer bound to instances to run once the final merge subtask had de-
the MapReduce processing pattern. Instead, we imple- termined the 2 · 108 smallest numbers, we changed
mented the sort/aggregation problem as a DAG and the communication channel between the final BigInte-
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


Stage 1
BigIntegerWriter (1)
ID: i-2E492091 In-memory channel
Type: m1.small
Network channel
BigIntegerAggregater (1) File channel

Stage 0
BigIntegerMerger (1)

Type: c1.xlarge Type: c1.xlarge

Type: c1.xlarge ... ... Type: c1.xlarge

BigIntegerSorter (126)
... ...
... 21 subtasks ... BigIntegerReader (126) ... 21 subtasks...
... ...
Fig. 6. The Execution Graph for Experiment 3 (DAG and Nephele)

gerMerger and BigIntegerAggregater subtask to a file following two MapReduce jobs, however, the allocated
channel. That way Nephele pushed the aggregation into instances are oversized: The second job, whose map and
the next Execution Stage and was able to deallocate the reduce phases range from point (d) to (f) and point (e) to
expensive instances. (g), respectively, can only utilize about one third of the
Finally, the fifth task, BigIntegerWriter, eventually re- available CPU capacity. The third job (running between
ceived the calculated average of the 2 · 108 integer num- point (g) and (h)) can only consume about 10 % of the
bers and wrote the value back to HDFS. overall resources.
The reason for Hadoop’s eventual poor instance uti-
lization is its assumption to run on a static compute
4.5 Results cluster. Once the MapReduce engine is started on a
Figure 7, Figure 8 and Figure 9 show the performance set of instances, no instance can be removed from that
results of our three experiment, respectively. All three set without the risk of losing important intermediate
plots illustrate the average instance utilization over time, results. As in this case, all six expensive instances must
i.e. the average utilization of all CPU cores in all in- be allocated throughout the entire experiment and un-
stances allocated for the job at the given point in time. necessarily contribute to the processing cost.
The utilization of each instance has been monitored Figure 8 shows the system utilization for executing
with the Unix command “top” and is broken down into the same MapReduce programs on top of Nephele. For
the amount of time the CPU cores spent running the the first Execution Stage, corresponding to the TeraSort
respective data processing framework (USR), the kernel map and reduce tasks, the overall resource utilization
and its processes (SYS), and the time waiting for I/O is comparable to the one of the Hadoop experiment.
to complete (WAIT). In order to illustrate the impact During the map phase (point (a) to (c)) and the reduce
of network communication, the plots additionally show phase (point (b) to (d)) all six “c1.xlarge” instances show
the average amount of IP traffic flowing between the an average utilization of about 80 %. However, after ap-
instances over time. proximately 42 minutes, Nephele starts transmitting the
We begin with discussing Experiment 1 (MapReduce sorted output stream of each of the 12 TeraSortReduce
and Hadoop): For the first MapReduce job, TeraSort, subtasks to the two instances which are scheduled to
Figure 7 shows a fair resource utilization. During the remain allocated for the upcoming Execution Stages. At
map (point (a) to (c)) and reduce phase (point (b) to (d)) the end of Stage 0 (point (d)), Nephele is aware that
the overall system utilization ranges from 60 to 80%. This four of the six “c1.xlarge” are no longer required for the
is reasonable since we configured Hadoop’s MapReduce upcoming computations and deallocates them.
engine to perform best for this kind of task. For the Since the four deallocated instances do no longer
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


q USR q USR
q SYS q SYS
q WAIT (b) q WAIT
Network traffic Network traffic
(a)

500

500
100

100
(c)

(b) (d)

450

450
(c) (g) (e) (f) (g) (h)

Average network traffic among instances [MBit/s]

(d)
(a)

400

400
80

80
(e)
Average instance utilization [%]

350

350
(f)
(h)

300

300
60

60
250

250
200

200
40

40
150

150
100

100
20

20
50

50
0

0

0

0
0 20 40 60 80 100 0 20 40 60 80 100

Time [minutes] Time [minutes]

Fig. 7. Results of Experiment 1: MapReduce and Hadoop Fig. 8. Results of Experiment 2: MapReduce and Nephele

contribute to the number of available CPU cores in the Initially, the BigIntegerReader subtasks begin to read
second stage, the remaining instances again match the their splits of the input data set and emit the created
computational demands of the first aggregation step. records to the BigIntegerSorter subtasks. At point (b) the
During the execution of the 12 AggregateMap subtasks first BigIntegerSorter subtasks switch from buffering the
(point (d) to (f)) and the four AggregateReduce subtasks incoming records to sorting them. Here, the advantage
(point (e) to (g)), the utilization of the allocated instances of Nephele’s ability to assign specific instance types
is about 80%. The same applies to the final aggregation to specific kinds of tasks becomes apparent: Since the
in the third Execution Stage(point (g) to (h)) which is entire sorting can be done in main memory, it only
only executed on one allocated “c1.xlarge” instance. takes several seconds. Three minutes later (c), the first
BigIntegerMerger subtasks start to receive the presorted
q USR
records and transmit them along the processing chain.
q SYS (e)
q WAIT Until the end of the sort phase, Nephele can fully
Network traffic (d)
exploit the power of the six allocated “c1.xlarge” in-
500
100

(c) stances. After that period the computational power is
450

(b)

no longer needed for the merge phase. From a cost per-
400

spective it is now desirable to deallocate the expensive
80

(a)
instances as soon as possible. However, since they hold
350

the presorted data sets, at least 20 GB of records must be
300
60

transferred to the inexpensive “m1.small” instance first.
250

Here we identified the network to be the bottleneck, so
much computational power remains unused during that
200
40

transfer phase (from (c) to (d)). In general, this transfer
150

penalty must be carefully considered when switching
100

between different instance types during job execution.
20

For the future we plan to integrate compression for file
50

and network channels as a means to trade CPU against
0

0

I/O load. Thereby, we hope to mitigate this drawback.
0 5 10 15 20 25 30 35
At point (d), the final BigIntegerMerger subtask has
Time [minutes]
emitted the 2 · 108 smallest integer records to the file
Fig. 9. Results of Experiment 3: DAG and Nephele channel and advises all preceding subtasks in the pro-
cessing chain to stop execution. All subtasks of the first
Finally, we want to discuss the results of the third stage have now been successfully completed. As a re-
experiment (DAG and Nephele) as depicted in Fig- sult, Nephele automatically deallocates the six instances
ure 9: At point (a) Nephele has successfully allocated of type “c1.xlarge” and continues the next Execution
all instances required to start the first Execution Stage. Stage with only one instance of type “m1.small” left. In
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


that stage, the BigIntegerAggregater subtask reads the in the execution environment. Pegasus uses DAGMan
2 · 108 smallest integer records from the file channel and and Condor-G [13] as its execution engine. As a result,
calculates their average. Again, since the six expensive different task can only exchange data via files.
“c1.xlarge” instances no longer contribute to the number Thao et al. introduced the Swift [30] system to re-
of available CPU cores in that period, the processing duce the management issues which occur when a job
power allocated from the cloud again fits the task to be involving numerous tasks has to be executed on a large,
completed. At point (e), after 33 minutes, Nephele has possibly unstructured, set of data. Building upon compo-
finished the entire processing job. nents like CoG Karajan [26], Falkon [21], and Globus [12],
Considering the short processing times of the pre- the authors present a scripting language which allows
sented tasks and the fact that most cloud providers offer to create mappings between logical and physical data
to lease an instance for at least one hour, we are aware structures and to conveniently assign tasks to these.
that Nephele’s savings in time and cost might appear The system our approach probably shares most simi-
marginal at first glance. However, we want to point out larities with is Dryad [14]. Dryad also runs DAG-based
that these savings grow by the size of the input data jobs and offers to connect the involved tasks through
set. Due to the size of our test cloud we were forced either file, network, or in-memory channels. However,
to restrict data set size to 100 GB. For larger data sets, it assumes an execution environment which consists of
more complex processing jobs become feasible, which a fixed set of homogeneous worker nodes. The Dryad
also promises more significant savings. scheduler is designed to distribute tasks across the
available compute nodes in a way that optimizes the
throughput of the overall cluster. It does not include the
5 R ELATED W ORK notion of processing cost for particular jobs.
In recent years a variety of systems to facilitate MTC has In terms of on-demand resource provising several
been developed. Although these systems typically share ¨
projects arose recently: Dornemann et al. [11] presented
common goals (e.g. to hide issues of parallelism or fault an approach to handle peak-load situations in BPEL
tolerance), they aim at different fields of application. workflows using Amazon EC2. Ramakrishnan et al. [22]
MapReduce [9] (or the open source version discussed how to provide a uniform resource abstraction
Hadoop [25]) is designed to run data analysis jobs over grid and cloud resources for scientific workflows.
on a large amount of data, which is expected to be Both projects rather aim at batch-driven workflows than
stored across a large set of share-nothing commodity the data-intensive, pipelined workflows Nephele focuses
servers. MapReduce is highlighted by its simplicity: on. The FOS project [28] recently presented an operating
Once a user has fit his program into the required map system for multicore and clouds which is also capable
and reduce pattern, the execution framework takes of on-demand VM allocation.
care of splitting the job into subtasks, distributing
and executing them. A single MapReduce job always 6 C ONCLUSION
consists of a distinct map and reduce program. However, In this paper we have discussed the challenges and
several systems have been introduced to coordinate the opportunities for efficient parallel data processing in
execution of a sequence of MapReduce jobs [19], [17]. cloud environments and presented Nephele, the first
MapReduce has been clearly designed for large static data processing framework to exploit the dynamic re-
clusters. Although it can deal with sporadic node fail- source provisioning offered by today’s IaaS clouds. We
ures, the available compute resources are essentially have described Nephele’s basic architecture and pre-
considered to be a fixed set of homogeneous machines. sented a performance comparison to the well-established
The Pegasus framework by Deelman et al. [10] has data processing framework Hadoop. The performance
been designed for mapping complex scientific workflows evaluation gives a first impression on how the ability to
onto grid systems. Similar to Nepehle, Pegasus lets its assign specific virtual machine types to specific tasks of a
users describe their jobs as a DAG with vertices repre- processing job, as well as the possibility to automatically
senting the tasks to be processed and edges representing allocate/deallocate virtual machines in the course of a
the dependencies between them. The created workflows job execution, can help to improve the overall resource
remain abstract until Pegasus creates the mapping be- utilization and, consequently, reduce the processing cost.
tween the given tasks and the concrete compute re- With a framework like Nephele at hand, there are a
sources available at runtime. The authors incorporate variety of open research issues, which we plan to address
interesting aspects like the scheduling horizon which for future work. In particular, we are interested in im-
determines at what point in time a task of the overall proving Nephele’s ability to adapt to resource overload
processing job should apply for a compute resource. This or underutilization during the job execution automati-
is related to the stage concept in Nephele. However, cally. Our current profiling approach builds a valuable
Nephele’s stage concept is designed to minimize the basis for this, however, at the moment the system still
number of allocated instances in the cloud and clearly requires a reasonable amount of user annotations.
focuses on reducing cost. In contrast, Pegasus’ schedul- In general, we think our work represents an important
ing horizon is used to deal with unexpected changes contribution to the growing field of Cloud computing
© 2011 IEEE
DOI 10.1109/TPDS.2011.65


services and points out exciting new opportunities in [20] I. Raicu, I. Foster, and Y. Zhao. Many-Task Computing for
the field of parallel data processing. Grids and Supercomputers. In Many-Task Computing on Grids and
Supercomputers, 2008. MTAGS 2008. Workshop on, pages 1–11, Nov.
2008.
[21] I. Raicu, Y. Zhao, C. Dumitrescu, I. Foster, and M. Wilde. Falkon:
R EFERENCES a Fast and Light-weight tasK executiON framework. In SC ’07:
Proceedings of the 2007 ACM/IEEE conference on Supercomputing,
[1] Amazon Web Services LLC. Amazon Elastic Compute Cloud pages 1–12, New York, NY, USA, 2007. ACM.
(Amazon EC2). http://aws.amazon.com/ec2/, 2009. [22] L. Ramakrishnan, C. Koelbel, Y.-S. Kee, R. Wolski, D. Nurmi,
[2] Amazon Web Services LLC. Amazon Elastic MapReduce. http: D. Gannon, G. Obertelli, A. YarKhan, A. Mandal, T. M. Huang,
//aws.amazon.com/elasticmapreduce/, 2009. K. Thyagaraja, and D. Zagorodnov. VGrADS: Enabling e-Science
[3] Amazon Web Services LLC. Amazon Simple Storage Service. http: Workflows on Grids and Clouds with Fault Tolerance. In SC
//aws.amazon.com/s3/, 2009. ’09: Proceedings of the Conference on High Performance Computing
[4] D. Battr´ , S. Ewen, F. Hueske, O. Kao, V. Markl, and D. Warneke.
e Networking, Storage and Analysis, pages 1–12, New York, NY, USA,
Nephele/PACTs: A Programming Model and Execution Frame- 2009. ACM.
work for Web-Scale Analytical Processing. In SoCC ’10: Proceed- [23] R. Russell. virtio: Towards a De-Facto Standard for Virtual I/O
ings of the ACM Symposium on Cloud Computing 2010, pages 119– Devices. SIGOPS Oper. Syst. Rev., 42(5):95–103, 2008.
130, New York, NY, USA, 2010. ACM. [24] M. Stillger, G. M. Lohman, V. Markl, and M. Kandil. LEO -
[5] R. Chaiken, B. Jenkins, P.-A. Larson, B. Ramsey, D. Shakib, DB2’s LEarning Optimizer. In VLDB ’01: Proceedings of the 27th
S. Weaver, and J. Zhou. SCOPE: Easy and Efficient Parallel International Conference on Very Large Data Bases, pages 19–28, San
Processing of Massive Data Sets. Proc. VLDB Endow., 1(2):1265– Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc.
1276, 2008. [25] The Apache Software Foundation. Welcome to Hadoop! http:
[6] H. chih Yang, A. Dasdan, R.-L. Hsiao, and D. S. Parker. Map- //hadoop.apache.org/, 2009.
Reduce-Merge: Simplified Relational Data Processing on Large [26] G. von Laszewski, M. Hategan, and D. Kodeboyina. Workflows
Clusters. In SIGMOD ’07: Proceedings of the 2007 ACM SIGMOD for e-Science Scientific Workflows for Grids. Springer, 2007.
international conference on Management of data, pages 1029–1040, [27] D. Warneke and O. Kao. Nephele: Efficient Parallel Data Process-
New York, NY, USA, 2007. ACM. ing in the Cloud. In MTAGS ’09: Proceedings of the 2nd Workshop
[7] M. Coates, R. Castro, R. Nowak, M. Gadhiok, R. King, and on Many-Task Computing on Grids and Supercomputers, pages 1–10,
Y. Tsang. Maximum Likelihood Network Topology Identification New York, NY, USA, 2009. ACM.
from Edge-Based Unicast Measurements. SIGMETRICS Perform. [28] D. Wentzlaff, C. G. III, N. Beckmann, K. Modzelewski, A. Belay,
Eval. Rev., 30(1):11–20, 2002. L. Youseff, J. Miller, and A. Agarwal. An Operating System for
[8] R. Davoli. VDE: Virtual Distributed Ethernet. Testbeds and Research Multicore and Clouds: Mechanisms and Implementation. In SoCC
Infrastructures for the Development of Networks & Communities, ’10: Proceedings of the ACM Symposium on Cloud Computing 2010,
International Conference on, 0:213–220, 2005. pages 3–14, New York, NY, USA, 2010. ACM.
[9] J. Dean and S. Ghemawat. MapReduce: Simplified Data Process- [29] T. White. Hadoop: The Definitive Guide. O’Reilly Media, 2009.
ing on Large Clusters. In OSDI’04: Proceedings of the 6th conference [30] Y. Zhao, M. Hategan, B. Clifford, I. Foster, G. von Laszewski,
on Symposium on Opearting Systems Design & Implementation, pages V. Nefedova, I. Raicu, T. Stef-Praun, and M. Wilde. Swift: Fast,
10–10, Berkeley, CA, USA, 2004. USENIX Association. Reliable, Loosely Coupled Parallel Computation. In Services, 2007
[10] E. Deelman, G. Singh, M.-H. Su, J. Blythe, Y. Gil, C. Kesselman, IEEE Congress on, pages 199–206, July 2007.
G. Mehta, K. Vahi, G. B. Berriman, J. Good, A. Laity, J. C. Jacob,
and D. S. Katz. Pegasus: A Framework for Mapping Complex
Scientific Workflows onto Distributed Systems. Sci. Program.,
13(3):219–237, 2005.
[11] T. Dornemann, E. Juhnke, and B. Freisleben. On-Demand Re-
source Provisioning for BPEL Workflows Using Amazon’s Elastic
Compute Cloud. In CCGRID ’09: Proceedings of the 2009 9th
IEEE/ACM International Symposium on Cluster Computing and the Daniel Warneke is a research assistant at the
Grid, pages 140–147, Washington, DC, USA, 2009. IEEE Computer Berlin University of Technology, Germany. He
Society. received his Diploma and BS degrees in com-
[12] I. Foster and C. Kesselman. Globus: A Metacomputing Infrastruc- puter science from the University of Pader-
ture Toolkit. Intl. Journal of Supercomputer Applications, 11(2):115– born, in 2008 and 2006, respectively. Daniel’s
128, 1997. research interests center around massively-
[13] J. Frey, T. Tannenbaum, M. Livny, I. Foster, and S. Tuecke. Condor- parallel, fault-tolerant data processing frame-
G: A Computation Management Agent for Multi-Institutional works on Infrastructure-as-a-Service platforms.
Grids. Cluster Computing, 5(3):237–246, 2002. Currently, he is working in the DFG-funded re-
[14] M. Isard, M. Budiu, Y. Yu, A. Birrell, and D. Fetterly. Dryad: Dis- search project Stratosphere.
tributed Data-Parallel Programs from Sequential Building Blocks.
In EuroSys ’07: Proceedings of the 2nd ACM SIGOPS/EuroSys Euro-
pean Conference on Computer Systems 2007, pages 59–72, New York,
NY, USA, 2007. ACM.
[15] A. Kivity. kvm: the Linux Virtual Machine Monitor. In OLS ’07:
The 2007 Ottawa Linux Symposium, pages 225–230, July 2007.
[16] D. Nurmi, R. Wolski, C. Grzegorczyk, G. Obertelli, S. Soman,
L. Youseff, and D. Zagorodnov. Eucalyptus: A Technical Report on Odej Kao is full professor at the Berlin University
an Elastic Utility Computing Architecture Linking Your Programs of Technology and director of the IT center tubIT.
to Useful Systems. Technical report, University of California, He received his PhD and his habilitation from
Santa Barbara, 2008. the Clausthal University of Technology. There-
[17] C. Olston, B. Reed, U. Srivastava, R. Kumar, and A. Tomkins. after, he moved to the University of Paderborn
Pig Latin: A Not-So-Foreign Language for Data Processing. In as an associated professor for operating and
SIGMOD ’08: Proceedings of the 2008 ACM SIGMOD international distributed systems. His research areas include
conference on Management of data, pages 1099–1110, New York, NY, Grid Computing, service level agreements and
USA, 2008. ACM. operation of complex IT systems. Odej is a
[18] O. O’Malley and A. C. Murthy. Winning a 60 Second Dash with member of many program committees and has
a Yellow Elephant. Technical report, Yahoo!, 2009. published more than 190 papers.
[19] R. Pike, S. Dorward, R. Griesemer, and S. Quinlan. Interpreting
the Data: Parallel Analysis with Sawzall. Sci. Program., 13(4):277–
298, 2005.

© 2011 IEEE
DOI 10.1109/TPDS.2011.65

Exploiting dynamic resource allocation for

More Related Content

What's hot (16)

Viewers also liked (6)

Similar to Exploiting dynamic resource allocation for (20)

More from ingenioustech (20)

Recently uploaded (20)

Exploiting dynamic resource allocation for