Scheduling in Bag-of-Task Grids: The PAUÁ Case.
ABSTRACT In this paper we discuss the difficulties involved in the scheduling of applications on computational grids. We highlight two main sources of difficulties: 1) the size of the grid rules out the possibility of using a centralized scheduler; 2) since resources are managed by different parties, the scheduler must consider several different policies. Thus, we argue that scheduling applications on a grid require the orchestration of several schedulers, with possibly conflicting goals. We discuss how we have addressed this issue in the context of PAUA, a grid for Bag-of-Tasks applications (i.e. parallel applications whose tasks are independent) that we are currently deploying throughout Brazil.
- SourceAvailable from: walfredo.dsc.ufcg.edu.brScientific Programming. 01/2002; 10:271-289.
- [show abstract] [hide abstract]
ABSTRACT: In heterogeneous and dynamic environments, efficient execution of parallel computations can require mappings of tasks to processors whose performance is both irregular (because of heterogeneity) and time-varying (because of dynamicity). While adaptive domain decomposition techniques have been used to address heterogeneous resource capabilities, temporal variations in those capabilities have seldom been considered. We propose a conservative scheduling policy that uses information about expected future variance in resource capabilities to produce more efficient data mapping decisions. We first present techniques, based on time series predictors that we developed in previous work, for predicting CPU load at some future time point, average CPU load for some future time interval, and variation of CPU load over some future time interval. We then present a family of stochastic scheduling algorithms that exploit such predictions of future availability and variability when making data mapping decisions. Finally, we describe experiments in which we apply our techniques to an astrophysics application. The results of these experiments demonstrate that conservative scheduling can produce execution times that are both significantly faster and less variable than other techniques.09/2003;
- Scalable Computing: Practice and Experience. 01/2000; 3.
Scheduling in Bag-of-Task Grids: The PAUÁ Case
Walfredo Cirne Francisco Brasileiro Lauro Costa
Daniel Paranhos Elizeu Santos-Neto Nazareno Andrade
Universidade Federal de Campina Grande
César De Rose Tiago Ferreto
Miranda Mowbray Roque Scheer João Jornada
In this paper we discuss the difficulties involved in the
scheduling of applications on computational grids. We
highlight two main sources of difficulties: firstly, the
size of the grid rules out the possibility of using a cen-
tralized scheduler; secondly, since resources are man-
aged by different parties, the scheduler must consider
several different policies. Thus, we argue that schedul-
ing applications on a grid require the orchestration of
several schedulers, with possibly conflicting goals. We
discuss how we have addressed this issue in the context
of PAUÁ, a grid for Bag-of-Tasks applications (i.e.
parallel applications whose tasks are independent) that
we are currently deploying throughout Brazil.
The use of computational grids as platform to exe-
cute parallel applications is a promising research area.
The possibility to allocate unprecedent amounts of
resources to a parallel application and to make it with
lower cost than traditional alternatives (based in paral-
lel supercomputers) is one of the main attractives in
grid computing. On the other hand, the grid character-
istics, such as high heterogeneity, complexity and wide
distribution (traversing multiple administrative do-
mains), create many new technical challenges.
In particular, the area of scheduling faces entirely
new challenges in grid computing. Traditional schedul-
ers (such as the operating system and the supercom-
puter scheduler) control all resources of interest. In a
grid, such a central control is not possible. First, the
grid is just too big for a single entity to control. Sec-
ond, the resources that comprise a grid are owned by
many different entities, rendering it administratively
unacceptable that a single entity controls all resources.
In a grid, a scheduler must strive for its traditional
goals (improve system and application performance),
while realizing that most of the system is not under its
control. In fact, most of the system will be under con-
trol of other schedulers. Therefore, a given scheduler
must interact with (or at least consider) other schedul-
ers in order to achieve its goals. In a way, the multiple
schedulers present in a grid form an ecology, where
individual schedulers compete and/or collaborate with
other schedulers, and the overall system behavior
emerges from the decisions made by all schedulers.
This paper discusses the scheduling ecology found in
PAUÁ, a 250-node grid that supports the execution of
Bag-of-Tasks applications. Bag-of-Tasks (BoT) appli-
cations are those parallel applications whose tasks are
independent of each other. Despite their simplicity,
BoT applications are used in a variety of scenarios,
including data mining, massive searches (such as key
breaking), parameter sweeps , simulations, fractal
calculations, computational biology , and computer
imaging  . Moreover, due to the independence
of their tasks, BoT applications can be successfully
executed over widely distributed computational grids,
as has been demonstrated by SETI@home . In fact,
one can argue that BoT applications are the applica-
tions most suited for computational grids, where com-
munication can easily become a bottleneck for tightly-
coupled parallel applications.
Focusing in BoT applications is interesting because
the problem is simplified, but remains useful and rele-
vant. The major simplification introduced by focusing
on BoT applications is that we do not need Quality-of-
Services guarantees. Since the tasks that compose a
BoT application are independent, having a task making
progress very slowly (or even stopping!) can be dealt
with no major problems. At worst, the task can be
The scheduling ecology found in PAUÁ was de-
signed to (i) respect site autonomy, (ii) cater for the
user’s priorities, (iii) enable multilateral collaboration
(in contrast to the much more common bilateral col-
laboration), (iv) support both dedicated and non-
dedicated resources, and (v) explicitly separate archi-
tectural components from implementation (thus easing
the addition of new schedulers to the ecology).
These design goals were achieved by separating the
grid scheduling concerns in three main aspects,
namely: (i) improve the performance of the application
over the whole grid, (ii) manage resources within a site
(i.e. a set of resources within a single administrative
domain), and (iii) gain access to resources throughout
the grid. Concern (i) is the responsibility of a job
scheduler, whereas concern (ii) is the responsibility of
a site scheduler. The grid has many job schedulers, as
well as many site schedulers. A user has rights of use
(i.e., an account) in one or more sites. Moreover, sites
can dynamically grant access to foreign users via a
peer-to-peer mechanism called network of favors, thus
addressing concern (iii).
Section 2 surveys the area and discusses related
work. Section 3 describes the experience in building
the PAUÁ’s community. We present PAUÁ’s schedul-
ing ecology in Section 4. Section 5 presents a set of
experiments that gauges the performance one can ex-
pect from PAUÁ. Section 6 contains our conclusions
and delineates future work.
2 Related Work
Grid computing is a very active area of research 
. Although it has started within High Performance
Computing, people have realized that Grid technology
could be used to deliver computational services on-
demand. This observation has brought about the merge
between Grid and Web Services technologies, as seen
in standards like OGSA/OGSI  and its successor
WSMF . These standards are currently being im-
plemented by both academia and industry. Most nota-
bly, these standards are being implemented by Globus
, maybe the project with greatest visibility in Grid
However, it is important to realize that WSMF-like
technologies do not address scheduling or resource
management directly. They rather provide the grid
building blocks, the common foundation on which
grids are built. Scheduling is thought to happen perva-
sively throughout the grids, with each service making
its own scheduling decisions (which may be delegated
to specialized services) . That is, the overall grid
scheduling is a result of the scheduling decisions made
by multiple autonomous (yet related) entities. In par-
ticular, the scheduling decisions made by a given ser-
vice s must take into account the quality of service
provided by the services invoked by s.
The idea that a single scheduler cannot deal with the
entire grid dates from the mid 1990s, with Berman et al
seminal work on application-level scheduling .
Since then, there has been a number of works on the
many aspects of scheduling in grids. These aspects
include, for example, coping with the d ynamicity of
grid resource availability (e.g.), the impact of large
data transfers (e.g.), coordination among many
schedulers to deliver a combined service (e.g.), and
virtualization as a way to ease scheduling (e.g.).
Closer to our work, there are scheduling efforts that
target BoT applications, such as APST  , Nim-
rod/G  and Condor  . In particular, APST
and Nimrod/G are similar to MyGrid, our job sched-
uler, in intent and architecture. However, they require
much more information than MyGrid for scheduling.
Moreover, they also differ from MyGrid in the assump-
tions about the application and the grid. APST targets
divisible workloads, whereas in MyGrid the user is the
responsible for breaking the application’s work into
tasks. Nimrod/G assumes that the user is going to pay
for resources and hence scheduling is based on a grid
economy model .
Condor was initially conceived for campus-wide
networks , but has been extended to run on grids
. Whereas MyGrid, APST and Nimrod/G are user-
centric schedulers, Condor is system-centric scheduler.
Condor is thus closer to OurGrid, our site scheduler.
The major difference between OurGrid and Condor is
that OurGrid was designed to encourage people to
donate their resources to the community (since re-
sources received are made proportional to resources
donated), whereas in Condor this issue is taken off-line
(e.g. altruism or administrative orders lead people into
a Condor pool).
Condor and OurGrid create grids on which resource
providers and resource consumers are roles played by
the same people. As an alternative, public computing
efforts suggest a more asymmetrical view, on which
many people voluntarily donate resources to a few
projects of great public appeal. Arguably, public com-
puting originated from the huge success achieved by
SETI@home , which has harvested close to
2,000,000 years of CPU so far . SETI@home
makes no distinction between the application itself
(search of extraterrestrial intelligence evidence in radio
signals) from its grid support. However, BOINC 
has been introduced as a SETI@home sequel, promis-
ing exactly such as separation. BOINC aims to create a
public computing infrastructure that can be used by
different applications. The Bayanihan project also aims
to create a public computing infrastructure and carries
a very interesting contribution on tolerating sabotage
(i.e. bogus volunteer results) .
3 The PAUÁ Community
PAUÁ, which means “everything” in Tupi-Guarani
(an ancient language spoken by native Brazilians), is
an initiative created by HP Brazil R&D to build a
countrywide Brazilian Grid. PAUÁ currently involves
11 different universities and research centers that col-
laborate with HP Brazil R&D in what we call the “HP
Brazil’s research ecosystem”. The goals of PAUÁ are
twofold. The first goal is to take advantage of a number
of computational resources available on the different
research centers as well as HP Brazil R&D itself, creat-
ing a wide, geographically distributed Grid along the
country. The second goal to foster grid research, so that
the solution currently being developed is constantly
improved based on its own usage and experience.
UFCG is responsible for the MyGrid and OurGrid as
well as research on independent auditing of SLAs
(Service Level Agreements) in Grids, integration with
supercomputers, among others. Instituto Atlântico
focuses on security aspects in the Grid. UNISINOS is
also focusing on security as well as management as-
pects. Instituto Eldorado is adding Windows support as
well as helping the community on the configuration
management and training. Hewlett Packard Brazil
R&D is doing research on idle cycle exploitation and
applications’ execution security (sandboxing). IPT/SP
is working on testing, applications development and
web services. CPAD/PUC-RS is performing research
on clusters’ integration so that cluster’s resources can
be used in a transparent manner. CAP/PUC-RS is de-
veloping Grid applications. LNCC is working on the
field of Bioinformatics applications. Finally, UNIFOR
and UNISANTOS are doing research on using grids to
perform data mining.
There are several challenges that arise when coping
with the decentralized administration of such geo-
graphically distributed community. Just to cite a few,
one must take into account the evolution of each re-
search being carried out, synchronize the correct time
each research institution joins the community, in-
creases or decreases the resources allocated to a given
institution, and plan the integration of a new piece of
technology into the community common software. To
cope with these challenges, the grid’s policies are man-
aged and defined by a general committee, which is
formed by research center’s representatives. The com-
mittee is responsible for establishing and defining a
flexible, dynamic and non-anarchical community, as
well as synchronizing and attuning all the different
activities being developed. The committee has regular
tele-conference meetings, which are used to track inte-
gration activities and define the next steps. Because of
the number of different people involved, besides hav-
ing regular tele-conference meeting, the committee
also meets face-to-face a couple of times a year.
4 Scheduling in PAUÁ
A grid such as PAUÁ poses many challenges for its
schedulers. The resources are widely spread, making it
very difficult to have an efficient global snapshot of the
grid. Also, there are multiple users and multiple re-
source owners, each with particular wishes and priori-
ties. This scenario creates the need for the system to
have multiple schedulers. We thus have designed and
implemented a set of schedulers that are collectively
responsible for the scheduling in PAUÁ. As discussed
before, these schedulers must respect the autonomy of
each site, considering the priorities associated to dif-
ferent users. Further, it must support both dedicated
and non-dedicated resources. Finally, their interaction
needs to be so that facilitates the addition of new
schedulers to the grid.
We achieved the above design goals by separating
the grid scheduling concerns in three main aspects. A
key aspect is the improvement of the performance of
the application over the whole grid. This is achieved by
a job scheduler that does so by following a very effi-
cient and lightweight approach, as will be explained
shortly. The next aspect is the definition of the concept
of a site, within which resources are managed follow-
ing a particular policy (i.e. a site comprises a set of
resources within a single administrative domain). A
site scheduler is in charge of imposing the site policy.
The final aspect is that of providing a way to gain
access to resources throughout the grid (i.e. resources
of a foreign site). This is the responsibility of a peer-to-
peer resource exchange network involving all site
The user submits a BoT job to a job scheduler, which
sends a request for resources to all sites the user has an
account on. Each of these sites is controlled by a site
scheduler, which allocates resources to the job sched-
uler in a best-effort basis. These resources may be local
resources (which are controlled by the site scheduler
itself) or foreign resources (which are obtained via the
network of favors). As the job scheduler begins to
receive resources from the site schedulers, it starts to
farm out the tasks that compose the application. The
goal of the job scheduler is to minimize the application
execution time. Note that resources are offered to the
job scheduler in a best-effort basis. That is, resources
may “disappear” at any time. As such, the job sched-
uler itself must guarantee that tasks finish by resubmit-
ting them whenever necessary.
In PAUÁ parlance, the peer-to-peer resource ex-
change network of favors is called OurGrid. Therefore,
the site scheduler is an OurGrid peer. The job sched-
uler is termed MyGrid.
4.1 Job Scheduler
Despite the simplicity of BoT applications, schedul-
ing BoT applications on grids is difficult. Grids intro-
duce two issues that complicate matters. First, efficient
schedulers depend on a lot of information about appli-
cation (such as estimated execution time) and resources
(processor speed, network topology, and so on), how-
ever this kind of information is typically difficult to
obtain . Second, since many important BoT appli-
cations are also data-intensive applications, consider-
ing data transfers is paramount to achieve good per-
formance. Thus, in order to achieve efficient schedules,
one must provide a coordinated data and computation
scheduling, which is a non-trivial task.
MyGrid’s first scheduler (Workqueue with Replica-
tion, or WQR) dealt only with the first issue. WQR
uses task replication to recover from bad task to ma-
chine allocations (which are inevitable, since it uses to
information). WQR performance is as good as tradi-
tional knowledge-based schedulers fed with perfect
information, at the cost of consuming more cycles .
However, WQR does not take data transfers into ac-
With version 2.0 of MyGrid, we released an alterna-
tive scheduler for MyGrid: Storage Affinity, which
does tackle both problems simultaneously. But note
that WQR is still available within MyGrid because it
does quite a good job with CPU-intensive BoT applica-
There are a few grid schedulers that take data trans-
fers into account in order to improve the performance
of the applications. Of those, the one that likely had
greater visibility is XSufferage . As its name sug-
gest, XSufferage is an extension of the Sufferage
scheduling heuristic, and therefore, is a knowledge-
In order to cope with lack of information about envi-
ronment and data placement concerns, we have devel-
oped a novel scheduling heuristic for data-intensive
BoT applications. This heuristic is named Storage
Affinity. The idea is to exploit data reutilization to
avoid unnecessary data transfers. The data reutilization
appears in two basic flavors: inter-job and inter-task.
The former arises when a job uses the data already
used by (or produced by) a job that executed previ-
ously, while the latter appears in applications whose
tasks share the same input data.
In order to take advantage of the data reutilization
and improve the performance of Data-Intensive BoT
applications, we introduce the storage affinity metric.
This metric determines how close to a site a given task
is. By how close we mean how many bytes of the task
input dataset are already stored at a specific site. Thus,
storage affinity of a task to a site is the number of bytes
within the task input dataset that are already stored in
We claim that information (data size and data loca-
tion) can be obtained a priori without difficulty and
loss of accuracy, unlike, for example, CPU and net-
work loads or the completion time of tasks. For in-
stance, this information can be obtained if a data server
at a particular site is able to answer requests about
which data elements it stores and how large is each
data element. Alternatively, an implementation of a
Storage Affinity heuristic can easily store a history of
previous data transfer operations containing the re-
Naturally, since Storage Affinity does not use dy-
namic information about the grid and the application,
inefficient task-to-processor assignments may occur.
In order to circumvent this problem, Storage Affinity
uses a task replication strategy similar to that used by
WQR . Replicas have a chance to be submitted to
faster processors than those processors assigned to the
original task, thus increasing the chance of the task
completion time to be decreased.
A total of 3,000 simulations were performed to inves-
tigate the efficiency of Storage Affinity against other
heuristics. Each simulation consisted of a sequence of
6 executions of the same job. These executions are
repeated for each of the 3 analyzed scheduling heuris-
tics (WQR, XSufferage and Storage Affinity). There-
fore, we have 18,000 execution time values for each
scheduling heuristic analyzed.
Table 1 presents a summary of the simulation results.
It is possible to note that, in average, Storage Affinity
and XSufferage achieve comparable performances.
The results show that both data-aware heuristics attain
much better performance than WQR. This is because
data transfer delays dominate the execution time of the
application, thus not taking them into account severely
hurts the performance of the application. In the case of
WQR, the execution of each task is always preceded
by a costly data transfer operation (as can be inferred
from the large bandwidth and small CPU waste). This
impairs any improvement that the replication strategy
of WQR could bring. On the other hand, the replication
strategy of Storage Affinity is able to cope with the
lack of dynamic information and yields a performance
very similar to that of XSufferage. The main inconven-
ience of XSufferage is the need for knowledge about
dynamic information, whereas the drawback of Storage
Affinity is the consumption of extra resources due to
its replication strategy (an average of 59% of extra
CPU cycles and a negligible amount of extra band-
width). Naturally, we do not report any wasting values
for XSufferage because this heuristic does not apply
any replication strategy.
Wasted Bandwidth (%)
Table 1 – Storage Affinity simulations results
From this result we can state that the Storage Af-
finity task replication strategy is a feasible tech-
nique to obviate the need for dynamic information
when scheduling data-intensive BoT applications,
although at the expenses of consuming more CPU.
We refer the reader to  for a complete per-
formance analysis of Storage Affinity.
4.2 Site Scheduler
Computational Grids are composed by resources
from several sites. Such resources can have differ-
ent processor architectures and use diverse operat-
ing systems. Consequently they may also differ in
performance, ranging from desktop machines to
supercomputers. In BoT grids we consider only
two types of resources: (i) space-shared parallel
machines, such as MPPs (Massive Parallel Proces-
sors) and clusters; and (ii) time-shared resources,
such as desktop-machines that can be accessed at
The access to site resources cannot be made
without a request. For example, in a MPP system,
the access to nodes cannot be made without a
request to the resource manager. This happens
because the resource manager decides when a job
will be executed and on which machine nodes it
will run. Strictly speaking, this is also true for
desktop machines, since there is no resource that
can be shared without the intervention of a re-
source scheduler. In this last case the operating
system will be the resource scheduler.
According to the Global Grid Forum Scheduling
Dictionary Working Group  there are two
types of resource schedulers involved at site level.
The Local Scheduler determines how the system
processes its job queue in the case of space shared
resources like a MPP or cluster. It is implemented
usually in the system resource manager. The Ma-
chine Scheduler is used when the resource is just
one machine (i.e. a desktop computer). This type
of scheduler uses some criteria to schedule jobs,
such as priority, length of time in the job queue
and available resources. It is implemented in the
operating system of the machine.
In this paper we are introducing a new type of
resource scheduler called Site Scheduler. It repre-
sents the site resources in the grid making them
available to the higher schedulers. Thus, access to
the schedulers described above (local and ma-
chine) must go through the site scheduler. The
main responsibilities for a site scheduler are: (i)
verification of access rights for grid jobs; (ii)
abstraction of site resource types for the grid; and
(iii) arbitration between site demand and grid
The verification of access rights is needed for
security reasons (i.e. to block grid users that are
not eligible to use site resources). Access rights
can also be used to impose limitations related to
the maximum number of allocated resources or the
exclusion of some specific resource types. Time
related restrictions are also possible like the exclu-
sion of grid accesses in peak hours. The resource
abstraction is an interesting feature since grid
users do not have to care about site resource types.
The negotiations the site manager has to do with
the local and machine schedulers to allocate the
site resources should be transparent to the grid.
Arbitration is also a key aspect since local users
and grid users will be competing for the same
resources. The site manager has to find a good
balance between them to maintain the external
interference to a reasonable level, and thus not
delay local users too much. A priority policy could
be used to guarantee a better response to special
Additional services may include caching for
tasks and executables (so the site scheduler would
function as a proxy), resource monitoring and
performance prediction of the whole site (to be
used in higher resource levels). Another interest-
ing issue is to consider the site scheduler the only
known IP of a site. In this case the grid would see
the site as one virtual resource that would have the
sum of a site’s available resources.
4.3 Network of Favors
To form a grid that shares resources between
multiple organizations, it is not only necessary to
have an infrastructure that allows the site schedul-
ers to use each other’s resources. It is also neces-
sary for the resource owners to make their re-
sources available to the grid. Although this is an
obvious statement, making this happen is not
straightforward. Our experience shows that mak-
ing people contribute their resources to the com-
munity is one of the hardest tasks in assembling a
grid. This experience is backed up by empirical
studies of several peer-to-peer resource sharing
communities, showing that in the absence of in-
centives for resource donation, most users only
consume resources from the system, donating
nothing back  .
To provide incentives for donating resources to
the grid, OurGrid implements a scheme called the
network of favors  . Each site offers to the
community access to its idle resources, expecting
to gain access to the idle resources of other par-
ticipants when its work exceeds its local capacity.
To motivate sites to share as many resources as
possible, the network of favors is designed to
promote fairness in the resource sharing; that is,
the more a site d onates to the community, the
more it should expect to receive from the commu-
In the network of favors, when one site con-
sumes resources owned by another site, that is
regarded as a favor paid by the resource owner to
the consumer. Every site in the system stores a
local balance of favors received minus favors
given for each other known site, based on its past
interactions with the other site. This balance is
updated on providing or consuming favors. When
there are conflicting requests for resources, the
resource owner prioritizes requests made by sites
with higher balances. The quantification of each
favor's value is done locally and independently –
negotiations and agreements are not used –and
affects only decisions of future resource alloca-
tions made by the two sites involved.
Sites that do not reciprocate favors satisfactorily
will over time be given lower priority by the
community. The non-retribution may happen for
several reasons, such as local resource failures, the
absence of resources at the site, or the use of the
desired resources locally or by other users at the
moment of the request. Free-riding sites may even
choose not to reciprocate favors. In any case, the
non-retribution of the favors gradually diminishes
the ability of the site to access the grid's resources.
This behavior is illustrated in Figure 1. This figure
shows the results of a simulation of a 100-site
community with different proportions f of free-
riders. It is possible to see that the fraction of the
community resources obtained by the free-riders
(epsilon) diminishes over time, tending to a very
Figure 1 – Resources obtained by free riders
We have also verified through simulations that
the amount of resources that a collaborator re-
ceives divided by the amount it donates (denoted
FR) is approximately 1. Figure 2 illustrates this for
a 100-site community in which the amount a site
donates is given by a uniform distribution U(1,19).
It is reasonable to assume that the cost of donating
a resource is smaller than the utility gained by
receiving it. Therefore, it is in the interest of sites
to donate as many resources as they can.
Figure 2 – Distribution of Favor Ratio
5 Running on PAUÁ
We have just started running BoT applications
over PAUÁ. Here we present the results attained
by a very simple experiment we have conducted.
The application used is a CPU-bound BoT appli-
cation that finds the divisors of a very large num-
ber. The experiment was conducted in a small
subset of PAUÁ, composed of four different sites.
Each site has a peer acting as a site scheduler
providing nodes to grid users. Table 3 shows the
sites used, their localizations, the maximum num-
ber of nodes available1, nodes configuration and
1 Only the idle resources were made available to the
grid. By idle we mean resources that are not being
used by local users. This shows how site schedulers
Site Location Max. Num-
ber of Nodes
CPAD Porto Alegre
Table 3 – Testbed configuration
Site schedulers communicate with the resource
schedulers within their sites in order to obtain
resources. The peers LSD, LCC and LabPetri have
to deal with only machine schedulers (the O.S. of
their machines). The CPAD site scheduler is more
complex; its major resources are controlled by
CRONO , a local scheduler for clusters. In this
way, this peer communicates with two types of
resource schedulers: the O.S. of desktop machines
and the CRONO local scheduler of its cluster
The experiments were executed as follows. A
MyGrid running on a desktop machine at the
CPAD site acts as the job scheduler. We per-
formed two sets of experiments with the environ-
ment described above in order to analyze the speed
up attained by the grid when compared to the
execution of the application on a standalone set-
ting. The first set of experiments was composed by
jobs with small tasks (1 minutes on a dedicated
Pentium III 733 MHz). The second set has tasks
that were 5.2 times longer than the tasks of the
first task. The first set obtained speedup ranging
from 6.2 to 11.1, with an average of 7.5. The peak
number of machine gained by the site scheduler to
its job scheduler was 33, in the fastest job exe-
cuted. The set of longer tasks had speedup ranging
from 14 to 31.3, with an average of 22.3. This set
reached a peak of 35 machines in utilization.
As it was expected, we obtained an improvement
in the application performance. However, they
also highlighted the overhead impact of the PAUÁ
structure. The set with small tasks does not present
a speedup as good as the second set due to the
latency to gain a grid machine (communication
can be configured in order to maintain performance of
local resources that are in use by local users.
2 The reason for the LSD site to have two peers is the
following: one acts as a site scheduler (lula), while
the other (robalo) acts as a relay. This is required be-
cause there is a connection lack due to the UFCG
firewall configuration. Only a peer (robalo) can make
connections to LCC, thus this machine was used as an
application relay of the grid to LCC resources.
between peers) and prepare an environment (e.g.
transfer the binary of task) to execute the tasks.
This implies that, at least for the moment, applica-
tion must have reasonable large granularity to
execute well in PAUÁ.
6 Conclusions and Future Work
We have presented a strategy to deal with the
scheduling of BoT application on large scale grids.
The strategy proposed is supported by a set of
schedulers divided in two distinct classes and a
peer-to-peer resource harnessing mechanism. Site
schedulers are responsible for providing grid re-
sources to job schedulers that, in turn, provide
efficient scheduling of BoT applications over the
available resources. We currently have two job
schedulers that use a replication mechanism to
achieve efficient scheduling without requiring any
information about the grid or the applications
being scheduled. Site schedulers ensure the im-
plementation of policies set by the resource owner.
They also try to find remote resources using a
peer-to-peer resource trading protocol (the net-
work of favors protocol). Our results show that the
ecology of schedulers work as intended, although
they require parallel applications of course grain to
fully benefit from the grid.
It is important to point out that the scheduling
problem is just one of the many problems that
need to be tackled in order to deploy a grid such as
PAUÁ. In particular, security issues are very chal-
lenging as well. We are currently working a vir-
tual-machine-based sandbox technology to address
some of these issues. Another clear avenue for
improvement is the relaxation of the BoT require-
ment we currently pose for the application. Our
strategy is to relax the application constraints
incrementally, supporting a broader class of appli-
cations at each step. Our next step will be what we
call workflow applications, i.e. parallel applica-
tions with tasks whose input comes from another
As described here, PAUÁ is currently deployed
in 4 of the 11 institutions that compose the grid.
The other 7 institutions currently use MyGrid
alone, i.e. they form a grid with (local and remote)
resources they have an account on. They are in-
stalling OurGrid and moving to the architecture
herein described in the next three months.
Although PAUÁ is (at least for now) a closed
grid, all the software d escribed in this paper is
open source. MyGrid and OurGrid are available at
www.ourgrid.org, whereas CRONO is available at
 D. Abramson, J. Giddy, L. Kotler. High Perform-
ance Parametric Modeling with Nimrod/G: Killer
Application for the Global Grid? IPDPS'2000, pp.
 D. Anderson, J. Cobb, E. Korpela. SETI@home: An
Experiment in Public-Resource Computing. Comm.
ACM, vol. 45, no. 11, pp 56-61, Nov. 2002.
 D. Anderson. Public Computing: Reconnecting
People to Science. Conference on Shared Knowl-
edge and the Web, Madrid, Spain, 2003.
 N. Andrade, W. Cirne, F. Brasileiro, P. Roisenberg.
OurGrid: An Approach to Easily Assemble Grids
with Equitable Resource Sharing. JSSPP’2003.
 N. Andrade, F. Brasileiro, W. Cirne, M. Mowbray.
Discouraging Free Riding in a Peer-to-Peer CPU-
Sharing Grid. HPDC’2004, June 2004.
 F. Berman et al. Application-Level Scheduling on
Distribute Heterogeneous Networks. Supercomput-
ing’96, Pittsburgh, 1996.
 F. Berman, G. Fox, T. Hey (Editors). Grid
Computing: Making The Global Infrastructure a
Reality. John Wiley & Sons, 2003.
 L. Burchard et al. The Virtual Resource Manager:
An Architecture for
Management. 4th International Symposium on
Cluster Computing and the Grid, 2004.
 R. Buyya, D. Abramson, J. Giddy. An Economy
Driven Resource Management Architecture for
Global Computational Power Grids. PDPTA 2000.
 H. Casanova et al. Heuristics for Scheduling Pa-
rameter Sweep Applications in Grid Environments.
9th HCW, pp 349-363, 2000.
 H. Casanova, J. Hayes, Y. Yang. Algorithms and
Software to Schedule and Deploy Independent
Tasks in Grid Environments. Workshop on Distrib-
uted Computing, Metacomputing, and Resource
Globalization. Aussois, France. December 2002.
 H. Casanova and F. Berman. Parameter Sweeps on
the Grid with APST. In , April 2003.
 K. Czajkowski et al. A Resource Management
Architecture for Metacomputing
JSSPP’1998, pp. 62-82, 1998.
 K. Czajkowski et al. From Open Grid Services
Infrastructure to WS-Resource Framework: Refac-
toring and Extension. http://www-106.ibm.com/
 W. Elwasif, J. Plank, R. Wolski. Data Staging
Effects in Wide Area Task Farming Applications.
IEEE International Symposium on Cluster Comput-
ing and the grid, pp. 122-129, May 2001,
 J. Frey et al. Condor-G: A Computation Manage-
ment Agent for Multi-Institutional Grids. 10th
HPDC, August 7-9, 2001.
 I. Foster and C. Kesselman. The Grid: Blueprint for
a New Computing Infrastructure. Morgan Kauf-
 Global Grid Forum. Grid Scheduling Dictionary of
Terms and Keywords. http://www.fz-juelich.de/
 Globus Web Site. http://www.globus.org
 N. Leibowitz, M. Ripeanu, A. Wierzbicki. Decon-
structing the KaZaA Network. 3rd IEEE Workshop
on Internet Applications, June 2003.
 M. Litzkow, M. Livny, and M. Mutka. Condor: A
Hunter of Idle Workstations. 8th ICDCS, 1988.
 M. Netto, C. De Rose. CRONO: A Configurable
and Easy to Maintain Resource Manager Opti-
mized for Small and Mid-Size GNU/Linux Cluster.
32nd ICPP'03, pp. 555-562, 2003.
 D. Paranhos, W. Cirne, F. Brasileiro. Trading
Cycles for Information: Using Replication to
Schedule Bag-of-Tasks Applications on Computa-
tional Grids. Euro-Par 2003.
 E. Santos-Neto, W. Cirne, F. Brasileiro, A. Lima.
Exploiting Replication and Data Reuse to Effi-
ciently Schedule Data-intensive Applications on
Grids. 10th JSSPP, June 2004.
 S. Saroiu, P. Gummadi, S. Gribble. A Measurement
Study of Peer-to-Peer File Sharing Systems.
MMCN’02, Jan 2002
 L. Sarmenta. Sabotage-Tolerance Mechanisms for
Volunteer Computing Systems. Future Generation
Computer Systems, 18:4, Elsevier, 2002.
 SETI@home Statistics Page. http://setiathome.
 S. Smallen et al. Combining Workstations and
Supercomputers to Support Grid Applications: The
Parallel Tomography Experience. HCW’2000.
 S. Smallen, H. Casanova, and F. Berman. Applying
Scheduling and Tuning to On-line Parallel Tomo-
graphy. Supercomputing’2001, Nov. 2001.
 W. Smith, I. Foster, V. Taylor. Scheduling with
Advanced Reservations. IPDPS’2000.
 J. R. Stiles et al. Monte Carlo Simulation of Neu-
romuscular Transmitter Release Using MCell, a
General Simulator of Cellular Physiological Proc-
esses. Comp. Neuroscience, pp. 279-284, 1998.
 S. Tuecke et al. Open Grid Services Infrastructure
(OGSI) Version 1.0. Global Grid Forum Draft Rec-
ommendation, 6/27/2003. http://www.globus.org/
 L. Yang, J.M. Schopf, I. Foster. Conservative
Scheduling: Using Predicted Variance to Improve
Scheduling Decisions in Dynamic Environments.
Supercomputing 2003, Nov. 2003.