AN EXTENDED MASTER WORKER MODEL FOR
A DESKTOP GRID COMPUTING PLATFORM (QADPZ)
Monica Vlădoiu
Departament of Informatics, PG University of Ploieşti Romania, Bd. Bucureşti, Nr. 39, Ploieşti, Romania
Zoran Constantinescu
ZealSoft Ltd., Str. Tg. Neamţ, Nr. 60, Bucureşti, Romania
Keywords: Grid computing, desktop grid computing, distributed and parallel computing, master worker paradigm.
Abstract: In this paper we first present briefly QADPZ, an open source platform for heterogeneous desktop grid
computing, which enables users from a local network (organization-wide) or Internet (volunteer computing)
to share their resources. Users of the system can submit compute-intensive applications to the system, which
are then automatically scheduled for execution. The scheduling is made based on the hardware and software
requirements of the application. Users can later monitor and control the execution of the applications. Each
application consists of one or more tasks. Applications can be independent, when the composing tasks do
not require any interaction, or parallel, when the tasks communicate with each other during the computation.
QADPZ uses a master worker-model that is improved with some refined capabilities: push of work units,
pipelining, sending more work-units at a time, adaptive number of workers, adaptive timeout interval for
work units, and use of multithreading, to be presented further in this paper. These improvements are meant
to increase the performance and efficiency of such applications.
1 INTRODUCTION
Usually, complex computational and visualization
algorithms require large amounts of computational
power. The computing power of a single desktop
computer is insufficient for running such complex
algorithms, and, usually, large parallel super-
computers or dedicated clusters are used for this job.
However, very high initial investments and
maintenance costs limit their availability. A more
convenient solution is based on the use of non-
dedicated desktop PCs in a Desktop Grid (DG)
computing environment. Harnessing idle CPU
cycles, storage space and other resources of
networked computers to work together on a
particularly computational intensive application, and
increasing power and communication bandwidth of
desktop computers provides for this solution.
In a DG system, the execution of an application
is orchestrated by a central scheduler node, which
distributes the tasks amongst the worker nodes and
awaits workers' results. It is important to note that an
application only finishes when all tasks have been
completed. The attractiveness of exploiting DGs is
further reinforced by the fact that costs are highly
distributed: every volunteer supports her resources
(hardware, power costs and internet connections)
while the benefited entity provides management
infrastructures, namely network bandwidth, servers
and management services, receiving in exchange a
massive and otherwise unaffordable computing
power. The usefulness of DG computing is not
limited to major high throughput public computing
projects. Many institutions, ranging from academics
to enterprises, hold vast number of desktop
machines and could benefit from exploiting the idle
cycles of their local machines.
We present briefly here QADPZ, an open source
platform for heterogeneous desktop grid computing,
which enables users from a local network (enterprise
DG) or Internet (volunteer computing) to share their
resources. The system allows a centralized
management and use of the computational power of
idle computers from a network of desktop
computers. Users of the system can submit compute-
intensive applications to the system, which are then
169
Vl
ˇ
adoiu M. and Constantinescu Z. (2008).
AN EXTENDED MASTER WORKER MODEL FOR A DESKTOP GRID COMPUTING PLATFORM (QADPZ).
In Proceedings of the Third International Conference on Software and Data Technologies - PL/DPS/KE, pages 169-174
DOI: 10.5220/0001882301690174
Copyright
c
SciTePress
automatically scheduled for execution. The
scheduling is made based on the application’s
hardware and software requirements. Users can later
monitor and control the execution of the
applications, each of them consisting of one or more
tasks. Applications can be independent, when the
composing tasks do not require any interaction, or
parallel, when the tasks communicate with each
other during the computation.
QADPZ has pioneered autonomic computing for
desktop grids and presents specific self-management
features: self-knowledge, self-configuration, self-
optimization and self-healing (Constantinescu,
2003). It is worth to mention that to the present the
QADPZ has over a thousand users who have
download it, and many of them use it for their daily
tasks (SOURCEFORGE, 2008).
2 MASTER WORKER
PARADIGM
The QADPZ framework is based on an extended
version of the master-worker paradigm. Some of the
disadvantages in using this model in a heterogeneous
and dynamic environment are briefly described here.
The traditional master-worker has the worker nodes
downloading small tasks from a central master node,
executing them, and then sending back the results to
the master. The master-worker computing paradigm
is built on the observation that many computational
problems can be broken into smaller pieces that can
be computed by one or more processes in parallel.
That is, the computations are fairly simple consisting
of a loop over a common, usually compute-
intensive, region of code. The size of this loop is
usually considered to be long.
In this model, a number of worker processes are
available, each can perform any one of the steps in a
particular computation. The computation is divided
into a set of mutually independent work units by a
master node. Worker nodes then execute these work
units in parallel. A worker repeatedly gets a work
unit from its master, carries it out and sends back the
result. The master keeps a record of all the work
units of the computation it is designed to perform.
As each work unit is completed by one of the
workers, the master records the result. Finally, when
all the work units have been completed, the master
produces the complete result. The program works in
the same way irrespective of the number of workers
available - the master just gives out a new work unit
to any worker who has completed the previous one.
Whereas the master worker model is easily
programmed to run on a single parallel platform,
running such a model for a single application across
distributed machines presents interesting challenges.
On a parallel platform, the processors are always
considered identical in performance. In a distributed
environment, and especially in a heterogeneous one,
processors usually have different types and
performance. This raises the problem of load
balancing of work-units between the workers in such
a way to minimize the total computing time of the
application. The ideal application is coarse-grain and
embarrassingly parallel. Granularity is defined as the
computation-to-communication ration, with coarse
grain applications involving small communication
time compared to computation time, and fine
grained application requiring much more time for
communication than computation.
Coarse-grain applications are ideal for DG
computing because most DG systems employ
commodity network links, which have limited
bandwidth and high latencies. Embarrassingly
parallel applications are those problems that easily
decompose into a collection of completely
independent tasks. Examples of such scientific
problems are: genetic and evolutionary algorithms,
Monte Carlo simulations, distributed web crawling,
image processing, image rendering.
In a heterogeneous environment, scheduling,
which includes both problem decomposition and
work-unit distribution (placement to workers), has a
dramatic effect on the program's performance. An
inappropriate decision regarding decomposition or
distribution can result in poor performance, due to
load imbalance. Effective scheduling in such
environments is a difficult problem.
This problem can be overcome by using
relatively simple heuristics, if appropriate
mechanisms are provided to the scheduler to
determine the complexity of the problem wrt.
computation and communication. This information
is then used to decompose the problem and schedule
the work-units in a manner that provides good load
balance, and thus good performance. The idea of
using dynamic creation of subtasks is also presented.
Subtasks are generated according to the
requirements of the problem, by taking into
consideration the available performance parameters
of the system (network bandwidth, latency, CPU
availability and performance).
ICSOFT 2008 - International Conference on Software and Data Technologies
170
3 IMPROVED MASTER
WORKER MODEL
We present in this section an improved version of
the master-worker model, which is based on an
algorithm with dynamic decomposition of the
problem and dynamic number of workers. The
improvements regard the performance of the original
model, by increasing the time workers are doing
computations, and decreasing the time used for
communication delays. This is achieved by using
different techniques, such as pipelining of the work-
units at the worker, redundant computation of the
last work-units to decrease the time to finish,
overlapped communication and computation at the
workers and the master, use of compression to
reduce the size of messages. We will describe in the
following subsections each of these techniques.
3.1 Pull vs. Push for Work-Units
In the original master-worker model, each time a
worker finishes a work-unit, it has to wait until it
receives the next work-unit for processing. If this
communication time is comparable with the time
needed for executing a work-unit, the efficiency of
the worker is reduced very much. The time intervals
used for communication and computation
(processing) are described in Figure 1.
The master-worker model uses the pull
technology, which is based on the request/response
paradigm. This is typically used to perform data
polling. The user (in our case the worker) is
requesting data from the publisher (in our case the
master). The user is the initiator of the transaction.
In contrast, a push technology relies on the
publish/subscribe/distribute paradigm. The user
subscribes once to the publisher, and the publisher
will initiate all further data transfers to the user. This
is better suited in certain situations.
Figure 1: Worker timeline in execution.
We first extend the master-worker model by
replacing the pull technology with the push
technology, as it is illustrated in Figure 2.
Figure 2: Pull vs. Push technology.
In this model, the worker doesn't send any more
requests for work-units. Instead, it first announces its
availability to the master when it starts, and the
master is responsible for sending further work-units.
The workers just wait for work-units, and process
them when received. At the end of each work-unit, it
sends back the results to the master. The master will
further send more work-units to the worker. This
moves all decisions about initiating work-units
transfers to the master, allowing a better control and
monitoring of the overall computation.
3.2 Pipelining of Work-Units
The worker efficiency increases if we reduce the
total time spent in waiting for communication. One
way to do that is to use work-units pipelining at the
worker, thus making sure that the worker has a new
work-unit available when it finishes the processing
of the current work-unit. Pipelining is achieved by
sending more that one work-unit to the workers, as
shown in Figure 3. Each worker will have at least
one more work-unit in addition to the one being
processed at that worker. This is done so that the
worker, after finishing a work-unit, will have ready
the next one for processing. In the beginning, the
master sends more than one work-units to the
worker, then after each received result, sends
another work-unit to be queued on the worker. The
worker does not need to wait again for a new work-
unit from the master after sending the result, the next
work-unit being already available for processing.
AN EXTENDED MASTER WORKER MODEL FOR A DESKTOP GRID COMPUTING PLATFORM (QADPZ)
171
Figure 3: Pipelining of worker tasks.
The immediate advantage of pipelining is that the
waiting time for a new work-unit is eliminated. This
is described in Figure 4. While the worker is
processing the next work-unit, a new work-unit is
sent by the master and is queued in the operating
system. When using non-blocking communication,
the waiting time for sending the result to the master
after finishing a computation can be also eliminated.
Figure 4: Worker timeline for unit pipeline.
Keeping one new work-unit available at the worker
seems to be enough to reduce the waiting time for
communication. However, there is a situation when
this is not adequate. It can happen that the execution
time of a work-unit is much shorter than the
communication time (consisting of sending back the
result and receiving the new work-unit). In this case,
the worker finishes the current work-unit, but the
new one is not yet received. Thus, a certain waiting
time is involved for receiving it (see Figure 5).
Figure 5: Unit pipeline - worst case.
If there are many work-units with short execution
times, than the overall waiting time can increase
significantly, reducing the efficiency of the worker.
The condition for this not to happen is the following:
execreswu
ttt
+
for the average
time values, or
(1)
(
)
−−
+
unitswork unitswork
execreswu
ttt
for the
individual
time values
This situation can be improved by pipelining more
than two work-units at the worker, thus using a
larger pipeline. The master starts by sending out a
number of work-units to the worker to fill the
pipeline. Each time a result is received back from
the worker, the master sends a new work-unit, thus
keeping the pipeline full. This algorithm works as
long as the average execution time for a work-unit is
larger than the average communication time for
sending a result and a new work-unit between the
worker and the master. If the communication time is
too large, the pipeline will eventually become empty
and the worker will need to wait for new work-units.
3.3 Sending More Work-Units at a
Time
To overcome this situation, the master needs to send
more than one work-units per each message. The
master starts by sending a message containing more
than one work-unit, and then keeps sending as long
as the pipeline is not full. Each time it receives a
result, it sends another work-unit, to compensate the
decreasing number of work-units from the pipe. If
the worker sends only one result per message back
to the master, and the master sends only one new
work-unit, then, eventually the pipeline will become
empty. In order to prevent that, the worker will need
to send back more results at a time.
We could consider, for example, that the number
of results per message is equal to the number of
work-units per message sent from the master. In this
case, all results from the work-units, which came in
one message, are sent back to the master the same
way in one message after all of them were
successfully computed.
Figure 6: More work-units per message.
ICSOFT 2008 - International Conference on Software and Data Technologies
172
This would solve the previous problem if the time to
send a larger message (more work-units) were much
smaller than the time to send the individual
messages (one work-unit). This is usually possible if
the data required to describe one work-unit is small
enough, so the messages are kept short. However, it
could still happen that communication time is larger
than the execution time, so that the worker will end
up waiting for new work-units. The condition for
this not to happen is the following:
nexecmresnwu
ttt
,,,
+ for the average time
values of multiple work-units per message
and execution.
(2)
3.4 Adaptive Number of Workers
As mentioned before, in a heterogeneous
environment based on idle desktop computers, the
total number of workers available could be changing
during the computation. New workers can register to
the master, and other can become temporarily
unavailable. The master controls the total number of
workers used for computation, since he is the one
sending out work-units to the workers. If necessary,
the master can choose not to use all the available
workers for computation, only a few of them. This
might be for different reasons, as described further.
In the master-worker model, the master can
become a bottleneck, especially when there are a lot
of workers, which connect to get work-units and
send results back. Overloading the master could
cause the bottleneck. Because the master has also to
do a small amount of processing each time when it
receives results from the workers, if too many
workers connect to the master, the processing
resource available might not be enough and the
request will be delayed. There is an upper limit on
the number of workers that can connect to the master
without overloading it. Finding out this number is
not easy and it depends on a variety of parameters
from the entire system: computational capabilities of
workers and master, communication delays, the
amount of processing involved for each results, etc.
Another bottleneck could be caused by too much
communication. Considering that there is enough
computational power on the master to serve a large
number of workers, it could happen that there are
too many messages exchanged between the master
and workers, thus communication delays can occur.
This might happen either because there are too many
messages per time unit, or because the amount of
data transferred is too high, exceeding thus the
available network bandwidth. On the contrary, if
there is too few workers used, then the total time of
the computation will be too large, not exploiting all
the resources available. This suggests that there is
some optimum for the number of workers, which
can increase the overall efficiency of the whole
computation, and reduce the time to complete it.
We define the overall efficiency of the
computation as being the ratio between the total
amount of time since the beginning of the
computation (t
total
) and the sum of execution times
for all completed work-units on all workers:
==
unitswork
exec
total
serial
parallel
system
t
t
t
t
E
(3)
We propose an adaptive algorithm to choose the
number of workers, based on performance measures
and estimates of execution times for work-units and
communication times for sending and receiving
messages. The number of workers is automatically
reduced if the efficiency of the computation
decreases. We employ a heuristic-based method that
uses historical data about the behavior of the
application. It dynamically collects statistical data
about the average execution times on each worker.
3.5 Use of Multithreading
The multithreaded programming paradigm allows
the programmer to indicate to the run-time system,
which portions of an application can occur
concurrently. Synchronization variables control the
access to shared resources and allow different
threads to coordinate during execution. The
paradigm has been successfully used to introduce
latency hiding in distributed systems or in a single
system where different components operate at
different speeds.
The paradigm of multithreading can provide
many benefits for the applications. In our situation,
it can provide good runtime concurrency, while
parallel programming techniques can be easier
implemented. The most interesting and probably
most important advantages are performance gains
and reduced resource consumption.
Operating system kernels supporting
multithreaded application perform thread switching
to keep the system reactive while waiting on slow
I/O services, including networks. In this way, the
system continues to perform useful work while the
network or other hardware is transmitting or
receiving information at a relatively slow rate.
Another benefit of multithreaded programming is
in the simplification of the application structure.
AN EXTENDED MASTER WORKER MODEL FOR A DESKTOP GRID COMPUTING PLATFORM (QADPZ)
173
Threads can be used to simplify the structure of
complex, server-type applications. Simple routines
can be written for each activity (thread), making
complex programs easier to design and code, and
more adaptive to a wide variation in user demands.
This has further advantages in the maintainability of
the application and future extensions.
The multithreaded paradigm can also improve
server responsiveness. Complex requests or slow
clients do not block other requests for service, the
overall throughput of the server being increased.
4 CONCLUSIONS
We have presented here our extended master-worker
model, which has been used in the QADPZ
development. Our model includes some refined
capabilities that are meant to increase the
performance and efficiency of the computation: push
of work units, pipelining, sending more work-units
at a time, adaptive number of workers, adaptive
timeout interval for work units, and use of
multithreading. Further measurements need to be
made in order to sustain our preliminary benchmark
tests that suggest significant improvements
regarding efficiency of the computation.
REFERNCES
Berman, F., et al., 2003. Grid computing: making the
global infrastructure a reality, J. Wiley, New York
BOINC. (2006) Open Source Software for Volunteer
Computing and Grid Computing (online) Available
from http://boinc.berkeley.edu, (Accessed 25
November 2007)
Constantinescu, Z., 2003 Towards an autonomic
distributed computing environment, in Proceedings of
14
th
International Workshop on Autonomic Computing
Systems, held in conjunction with 14
th
Int. Conf. on
Database and Expert Systems Applications DEXA
2003, pp. 694-698, Prague, Czech Republic
Constantinescu, Z. & Petrovic, P., 2002, Q2ADPZ* an
open source, multi-platform system for distributed
computing. ACM Crossroads, 9, pp. 13-20.
Constantinescu, Z., 2008, A desktop grid computing
approach for Scientific Computing and Visualization,
PhD Thesis, Norwegian University of Science and
Technology, Trondheim, Norway
Cummings, M. P., (2007), Grid Computing (online)
Available from http://serine.umiacs.umd.edu/
research/grid.php (Accessed 25 March 2008)
David, P. A., et. al., 2002 SETI@home: an experiment in
public-resource computing. Communications, ACM,
45, pp. 56-61.
Distributed.Net (2008) (online) Available from
http://distributed.net, (Accessed 5 March 2008).
Distributedcomputing.Info (2008) (online) Available from
http://distributedcomputing.info, (Accessed 5 March
2008)
Foster, I. & Kesselman, C., 1999. The grid: blueprint for a
new computing infrastructure, San Francisco, Morgan
Kaufmann Publishers.
Foster, I. & Kesselman, C., 2004. The grid: blueprint for a
new computing infrastructure, Amsterdam, Boston,
Morgan Kaufmann.
Garg, V. K., 1996. Principles of distributed systems,
Boston, Kluwer Academic Publishers.
Garg, V. K., 2002. Elements of distributed computing,
New York, Wiley-Interscience.
Globus (2007) Globus (online) Available from http://
www.globus.org, (Accessed 15 March 2008)
Juhasz Z., Kacsuk P., Kranzlmuller D., 2004, Distributed
and Parallel Systems: Cluster and Grid Computing,
New York, Springer
Karniadakis, G. & Kirby, R. M., 2003. Parallel scientific
computing in C++ and MPI: a seamless approach to
parallel algorithms and their implementation, NY,
Cambridge University Press.
Leopold, C., 2001. Parallel and distributed computing: a
survey of models, paradigms, and approaches, New
York, Wiley.
Mustafee, N. & Taylor, S. J. E., 2006. Using a desktop
grid to support simulation modelling, in Proceedings
of 28th International Conference on Information
Technology Interfaces (ITI 2006), Dubrovnik, Croatia,
pp. 557 - 562
QADPZ (2008) (online) Available from http://
qadpz.sourceforge.net. (Acc. 1 April 2008).
SETI@HOME (2003) (online) Available from http://
setiathome.ssl.berkeley.edu, (Acc5 May 2003)
SOURCEFORGE (2008) (online) Available from http://
sourceforge.net, Acc 1 April 2008
Sunderam, V. S., 1990 PVM: a framework for parallel
distributed computing. Concurrency: Practice.
Experience, Vol. 2, pp 315-339.
Vahid, G., Lionel, C. B., Yvan, L., 2006 Traffic-aware
stress testing of distributed systems based on UML
models, in Proceeding of the 28th International
Conference on Software engineering, Shanghai, China
Zomaya, A. Y., 1996
Parallel and distributed computing
handbook, New York, McGraw-Hill.
ICSOFT 2008 - International Conference on Software and Data Technologies
174