execution management and efficient resource provisioning ...doi: 10.1016/j.procs.2015.05.412...

doi: 10.1016/j.procs.2015.05.412

Execution management and efficient resource provisioning

for flood decision support

Bartosz Balis1, Marek Kasztelnik2, Maciej Malawski1, Piotr Nowakowski2,Bartosz Wilk2, Maciej Pawlik2, and Marian Bubak1

1 AGH University of Science and Technology, Department of Computer Science, Krakow, Poland2 AGH University of Science and Technology, ACC CYFRONET, Krakow, Poland

AbstractWe present a resource provisioning and execution management solution for a flood decision sup-port system. The system, developed within the ISMOP project, features an urgent computingscenario in which flood threat assessment for large sections of levees is requested within a spec-ified deadline. Unlike typical decision support systems which utilize heavyweight simulationsin order to predict the possible course of an emergency, in ISMOP we employ an alternativeapproach based on the ‘scenario identification’ method. We show that this approach is a partic-ularly good fit for the resource provisioning model of IaaS Clouds. We describe the architectureof the ISMOP decision support system, focusing on the urgent computing scenario and itsformal resource provisioning model. Preliminary results of experiments performed in order tocalibrate and validate the model indicate that the model fits experimental data.

Keywords: Flood decision support, flood threat assessment, urgent computing, resource provisioning,

execution management, cloud computing, scientific workflows

1 Introduction

A typical flood scenario in the Malopolska region of Poland is one wherein many kilometers oflevees are exposed to a passing flood wave which can last for up to several weeks. For example,during the flood of 2010 there were two major waves, from 14 May to 3 June and from 4 June to2 July, affecting 61 out of 182 boroughs of the region [23]. A high water level lasting for manydays can ultimately lead to the loss of stability of the levee structure and a breach. Onlinemonitoring of levees with in-situ sensors can help predict such events, make timely decisions,and initiate appropriate actions [11].

In this paper we present a solution for execution management and resource provisioning fora flood decision support system. Although currently levees are not instrumented with sensorson a massive scale even in countries most affected by floods, real-time monitoring of flooddefences has been rapidly gaining interest over the recent years. Consequently, we proposeand evaluate a solution which can scale to large areas of levees monitored with sensors. We

Procedia Computer Science

Volume 51, 2015, Pages 2377–2386

ICCS 2015 International Conference On Computational Science

Selection and peer-review under responsibility of the Scientific Programme Committee of ICCS 2015c© The Authors. Published by Elsevier B.V.

2377

http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.05.412&domain=pdf

http://crossmark.crossref.org/dialog/?doi=10.1016/j.procs.2015.05.412&domain=pdf

focus on an urgent computing scenario wherein the user requests the computation of flood threatassessment for a large area of levees in a specified deadline. The algorithm for threat assessmentis based on scenario identification, an approach alternative to on-demand simulations typicallyused in decision support systems. We argue that cloud computing is a perfect solution for suchan approach, and in particular we consider whether public clouds could be a good choice fora computing infrastructure of a flood decision support system.

This work is carried out within the ISMOP project: a computerized levee monitoring anddecision support system.1 Research conducted in ISMOP spans various aspects including theconstruction of an artificial levee [15], design of wireless sensors for levee monitoring, develop-ment of a sensor communication infrastructure [21], levee modeling and simulation [17], anda decision support system whose aspects are the subject of this paper.

The paper is organized as follows. Section 2 overviews related work. Section 3 presents theISMOP flood decision support system. In Section 4 the urgent computing scenario is described.Section 5 presents preliminary experimental results. Finally, Section 6 concludes the paper.

2 Related work

Traditionally scientific computations have utilized HPC or Grid computing infrastructures [18],both of which are also used for urgent computing systems. However, resource management inthese infrastructures, oriented towards batch processing, involves job queueing or even resourcebrokering (Grids). As a result, on-demand access to computing resources, crucial for urgentjobs, calls for complex job prioritization and preemption mechanisms. For example, SPRUCEintroduced the mechanism of right-of-way tokens for this purpose [4]. Another approach isbased on modifications to the scheduling system in order to ensure good turnaround times forurgent jobs [24].

Right from their inception computational clouds have been studied as an alternative com-puting infrastructure for e-Science applications. With a fundamentally different resource man-agement philosophy based on-demand resource provisioning, clouds are seen as particularlyuseful for applications where the results of computations must be obtained within a specifiedtime, such as simulation-based urgent clinical decision-making [20].

In our previous work we presented an early warning system factory, the Common InformationSpace (CIS) [2], which formed the basis for the UrbanFlood flood early warning system [11]. InCIS we utilized cloud services for on-demand deployment of early warning system instances andtheir autoscaling in emergency situations [1]. However, we have only been focusing on elasticresource provisioning in order to address variable resource demands, but not strictly on urgentcomputing scenarios.

In [10] the authors describe a model-based approach to management of heterogeneous com-puting resources for urgent computing scenarios. The solution is based on the CLAVIRE plat-form [22]. Knyazkov et al. propose a workflow-based infrastructure for urgent computing [8]which supports interactive capabilities important in some decision support systems e.g. thoseinvolving crowd management. Interactivity enables, amongst others, computational steeringwhich makes it possible to interactively change the parameters of the underlying simulation.

1http://www.ismop.edu.pl

Execution management and efficient resource provisioning for flood decision support Balis, et. al

2378

3 ISMOP Flood Decision Support System

The overview of the ISMOP Flood Decision Support Platform is shown in Fig. 1. The platformoffers two main functions to the end user, the first being the levee monitoring and flood decisionsupport system, which is the focus of this paper. Besides the decision support system, ISMOPalso provides a virtual laboratory for conducting scientific experiments related to monitoringand controlled flooding of the artificial levee built within the framework of the project. Sensordata is delivered via the ISMOP telemetry system [21, 5] which ensures energy-aware dataacquisition with appropriate SLA guarantees [9] in an emergency.

Figure 1: ISMOP Flood Decision Support Platform overview.

The decision support system involves various data analyses performed automatically or on-demand, including time series analysis of sensor measurements using statistical methods (e.g.anomaly detection), prediction of the levee state, and flood threat assessment which can beperformed in the urgent computing mode.

The flood threat assessment computation is based on the scenario identification method,sketched out in Fig. 2. We assume that the levees are logically partitioned into small (e.g. 10-meter long) sections. For each such section there exists a database of pre-computed scenariossimulating the levee behavior for various initial and boundary conditions. The basis of the floodthreat assessment computation is the comparison of the current levee state, indicated by recentsensor measurements, with the database of pre-simulated scenarios, and finding one or severalscenarios which constitute the best fit. If the scenarios identified as best matches predict leveebreach within a certain time, the flood threat level for the given levee section is increased. Thecalculations involved in the computation of the threat level boil down to comparing time seriesdata sets. More details on the method can be found in [6].

Typical decision support systems involve heavyweight simulations in order to predict thedevelopment of an emergency situation. The scenario identification is an alternative approachwhich carries a number of benefits from the perspective of an urgent computing platform.


2379

Figure 2: Flood threat assessment computation based on scenario identification.

1. Demand for computing resources. Simulations have much higher resource demandsin comparison to scenario identification. Deadline-driven scheduling of simulations is also moredifficult and less accurate because their performance model – required for estimation of thenecessary quantity of computing resources – is much more complex.

2. Fault tolerance. Some simulations used in decision making may consume hundreds ofthousands of cores. At this scale, traditional approaches to fault tolerance (checkpointing or re-run) – especially in the case of deadline-driven computations – are regarded as ineffective [12].Sometimes complex fault-tolerant algorithms are devised as a remedy. In contrast, scenarioidentification computations are composed of a huge number of fine-grained, independent jobswhich are easy to rerun if needed.

3. Suitability of computing infrastructures. Simulations not only require hugeamounts of computing resources, but can also be sensitive to performance characteristics ofother hardware components, such as interconnects. Consequently, HPC infrastructures havebeen considered as the best choice for urgent computing [12]. However, as indicated earlier,HPC infrastructures are also challenging in terms of usability, on-demand access and QoS guar-antees required for urgent jobs. Scenario identification, on the other hand, is well-suited forcloud infrastructures.

4 Flood threat assessment: urgent computing use case

The flood threat assessment scenario performed in the urgent computing mode is particularlychallenging as it requires the allocation of a certain amount of computing resources in orderto meet the prescribed deadline. In this section we discuss two important components of oursolution: the architecture of the system and the resource provisioning model.

4.1 Architecture

Flood threat assessment based on scenario identification can be performed as an urgent com-putation. The parameters of such a computation are as follows:

1. Target levee sections. Each section results in an independent computational job.


2380

2. Sensor data time window size which specifies the extent of the time series representingcurrent sensor readings to be compared with the simulated time series (scenarios). Theduration of a simulated scenario spans around 10-14 days, while the sensor data timewindow may vary but will typically span 1-3 days. The longer the time window, the moreaccurate the flood threat assessment; however this also results in heavier computationaljobs.

3. Deadline. The final time by which the urgent computation should be completed.

The architecture implementing this scenario is shown in Fig. 3.Starting the computation. The user interacts with the graphical interface and decides

which area should be immediately assessed against the flood threat. The user also specifies thedeadline in which results must be returned.

Workflow generation. Based on the user’s selection, the system first determines whichlevee sections are to be assessed. Subsequently, it generates a workflow describing the compu-tational jobs that need to be performed, along with their data dependencies.

Figure 3: Implementation of the urgent computing scenario for flood threat assessment.

Calculating an optimized execution plan. The choice of the workflow model of com-putation allows us to employ scheduling algorithms in order to calculate the deployment andexecution plan for the workflow suitable for the specified deadline [13, 14]. The performancemodel used in the current prototype is described in Section 4.2.

Provisioning of resources. Next, the system allocates the necessary number of VirtualMachine instances and may potentially adjust this number as the workflow execution progresses,in accordance with the deployment plan. The initial provisioning of resources as well as anyfurther adjustments are done by the Atmosphere platform2.

2https://github.com/dice-cyfronet/atmosphere


2381

Workflow execution. The workflow is orchestrated by the HyperFlow workflow manage-ment system [3]. HyperFlow determines which jobs are ready to be executed (their dependenciesare fulfilled) and sends the appropriate job descriptions to a job queue. Jobs are fetched fromthe queue by Cloud Executors deployed on Virtual Machines. The Cloud Executors are alsoresponsible for fetching input data required by the job directly to the Virtual Machine (nodata is transferred through the workflow engine). The execution plan takes into account datalocality: if multiple jobs utilize the same data sets, they will be clustered and assigned to thesame workers in order to avoid unnecessary data transfers.

Workflow structure. The simplified structure of the workflow is shown in Fig. 4. EachComputeScenarioRanks process of the workflow computes the ranking of best matching sce-narios for a single levee section. The best match is computed by comparing the current sensorreadings from a levee section (represented by a RealData input) with those simulated in thescenarios (the SimulatedScenarios input). The ranking of scenarios is used to compute thethreat level for each levee section. If the scenarios selected as best matches predict a highlikelihood of levee breach, the threat level for a given section is increased accordingly.

Figure 4: The flood threat assessment workflow (sample levee with 200 sections). RealData

represents recent sensor measurements from a given levee section. Simulated scenarios isa data set of pre-simulated scenarios for that section. Multiple adjacent sections share thesame set of simulated scenarios.

4.2 Resource provisioning model

The goal of the provisioning platform is to ensure that the capacity of resources is sufficient tocomplete the whole workflow within the required deadline. In order to estimate the requiredamount of resources in terms of virtual machine instances, a performance model of the appli-cation needs to be developed and calibrated experimentally. Here we describe the performancemodel based on a bag-of-tasks [16, 7] approximation of workflows. The performance model isa function T = f(v, d, s, ...) that estimates the time T of execution of the application based onthe infrastructure and application parameters, such as number of VMs v, problem size d (inour case it is the size of the window in days), number of tasks s (representing the number ofsections) or other parameters.

In our bag-of-task performance model, we assume that all the tasks in the workflow areuniform and can be executed independently. This approximation can be applied to workflowsin which we can distinguish a set comprising a large number of dominating tasks. In the floodthreat assessment workflow the ComputeScenarioRanks tasks are dominating. Therefore, the


2382

approximation of T should include the following terms: computation time expressed as the sumof task execution times, tk, divided by the number of VMs; overheads that are proportionalto the number of VMs, v; and some constant term representing a possible constant overhead,such as VM provisioning delay. This model can be further simplified by assuming that taskexecution times are identical and equal to t. Moreover, we assume that t is a linear functionof the window size d. This results in Eq. 1, where a, b and c are parameters of the model thathave to be determined experimentally. This equation can be solved, yielding an estimate of thenumber of processors, if deadline T is given:

T =

∑sk=1 tkv

+ b ∗ v + c =s ∗ tv

+ b ∗ v + c = a ∗ s ∗ dv

+ b ∗ v + c (1)

The model can be further extended by treating task execution times as random variableswith an empirically derived distributions, allowing us to estimate the number of resourcesrequired to meet the deadline with a desired probability. Other models considered for futurework may include multi-level workflows [14].

5 Preliminary experiments

We have performed experimental evaluation of the urgent flood threat assessment application.The goals of these experiments are twofold: first, to asses the performance of the system andits overheads, and second, to obtain the data needed for the calibration of the applicationperformance model.

Preliminary experiments have been performed on a private cloud infrastructure consistingof a single node with 8 cores, equipped with Intel Xeon E5-2650 processors. The runtimecomponents of the HyperFlow engine were deployed on small virtual machines (single virtualcore, 512MB RAM). The data for simulated scenarios (244MB total) was generated and pre-installed on disks attached to each VM.

●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●

●●●●●●●●●●●●●●●●●●●●●●

●●●●

●

●●●●●

●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●

●

●●●

●

●

●●●●

●

●

●

●●

●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●

●

●●●●●●●●

●

●

1 4 8 16

2060

100

140

# VMs

Tim

e [s

]

(a) 1024 sections, 1 day

●

●●●●●●●●●●●●●●●●

●

●

●●●

●

●

●●●●●

●

●

●

●●

●●

●●●

●●●●●

●

●● ●

●

●●●

●

●

●

●●●●●●●

●●●●

●

●

●●●

●●●●●

●

●●●●●●

●●●●●

●

●●●●●

●●

●●

●

●

●● ●

●●

●●

●●●

●●

●●

●

●

●

●

●●●●

●

●

●●●

●

●●●●

●

●●●

1 2 3 4 5 6 7

2060

100

140

# Days

Tim

e [s

]

(b) 128 sections, 16 vms

●

●

1 4 8 16

2040

6080

120

# VMs

Tim

e [s

]

(c) warmup tasks only: 1024 sec-tions, 1 day

Figure 5: Execution time of ComputeScenarioRanks tasks. Outliers represent “warmup” tasksand the linear function is used to separate them as shown in (a) and (b). In (c) the line showsthe linear fit to the data.

The preliminary measurements of execution times of ComputeScenarioRanks tasks, shownin Fig. 5, indicate that the average execution time is proportional to the time window sizeexpressed in days (see section 4.1 and Fig. 5b). The outliers shown in the boxplots correspondto first tasks executed on each VM right after boot (“warmup” tasks). These tasks have


2383

considerably longer run times that are proportional to the number of VMs as shown in Fig. 5c.This can be attributed to the combined effects of the system warm-up, which includes theoverhead induced by other VMs that are still booting up, initial access to files (which can becached for subsequent reads), and other initialization overheads.

We used the data from our measurements to calibrate our performance model using non-linear regression and equation 1. This yielded the following coefficients: a = 6.53, b = 9.41, andc = 31.71. Fig. 6 shows how the total computing time of the workflow (including VM startuptimes) depends on the number of VMs. The values predicted by our model are also shown (redlines). It can be seen that – owing to linear dependencies on parameters such as the numberof days and sections – the model matches experimental data with good accuracy. The onlydiscrepancy is observed for 16 VMs and this can be attributed to the fact that in this case 16VMs were running on 8 physical cores, hence the task execution time was longer. Nevertheless,good speedup can be observed: for 1024 sections the execution time was reduced from 116 to13 minutes, which gives a speedup of 8.6.

● ●●●●●●●●

●

●

●

●

020

0040

0060

00

# VMs

Tota

l tim

e [s

]

1 4 8 16

Figure 6: Total computing time as a function ofthe number of VMs: comparison of measuredvalues (points) and predictions of our model(lines) for 128 and 1024 sections.

These results suggest that the overhead ofthe proposed system based on workflows andclouds is negligible and outweighed by the ad-vantages offered by using clouds. It shouldalso be noted that results depend on the se-lected cloud infrastructure and VM flavor (in-stance type), so that the performance modelneeds to be calibrated against all availabletypes of resources. This, fortunately, is nota problem, since the flexible architecture ofexecution and provisioning platforms allowsus to run benchmarks on other clouds with-out much effort. Our future work will addressthis aspect.

6 Conclusion

Clouds are increasingly used for scientificcomputations and they can also provide aneffective computing infrastructure for urgentcomputing systems whose workload can be characterized as spiky: most of the time its re-source demands are low, with occasional bursts which occur in rare emergency situations. Inthe ISMOP flood decision support system we propose an alternative approach to flood threatassessment based on scenario identification. From the performance and workload characteriza-tion perspective scenario identification comprises many independent jobs which is considerablydifferent from on-demand simulations, typically used in decision support systems leveragingurgent computations. Consequently, scenario identification matches the resource provisioningmodel of IaaS clouds. Furthermore, scenario identification as a computational job can be de-scribed by a relatively simple resource provisioning (performance) model which can be easilycalibrated for different cloud infrastructures using simple benchmarks.

While in this paper we have focused on IaaS clouds from the resource provisioning modelperspective, we strongly believe that public cloud services can benefit flood decision supportsystems in various aspects other than the provisioning model and performance, including:


2384

• Reliability: public cloud services are specifically designed to support systems with highavailability demands, such as Web pages and online services maintained by the largestnews or media content providers. According to [19], Amazon had only five major outagesin the years 2010-2013, of which only one (in April 2011) lasted more than 6 hours.

• Safety: In the aftermath of a serious natural disaster any local computing infrastructure,and hence decision support systems which depend on it, may become inoperable, as wasthe case after Japan’s March 2011 earthquake [12]. Public clouds not only serve as anemergency computing infrastructure, but they can also increase data safety simply byproviding a reliable storage infrastructure where important – but not sensitive – datacan be backed up. In the scenario identification approach pre-simulated scenarios are anexcellent example of such data.

• Cost-effectiveness: Due to the ‘spiky’ behavior of decision support systems (in relationto resource demands), the cost of maintaining a dedicated infrastructure is certain tooutweigh the costs of occasional heavy computations performed in the cloud. Even thoughmonitoring of natural phenomena in a decision support system must be continuous, itsnormal day-to-day operation can be handled by a relatively small, low-cost on-premisesinfrastructure.

Future work involves, among others, further enhancement and validation of the resourceprovisioning model, investigation of alternative performance models (e.g. dynamic bag of tasks),and experimental benchmarks on public cloud infrastructures.

Acknowledgments

This work is partially supported by the National Centre for Research and Development(NCBiR), Poland, project PBS1/B9/18/2013; and by AGH University of Science and Tech-nology, Faculty of Computer Science, Electronics and Telecommunications, statutory projectno. 11.11.230.124.

References

[1] B. Balis, T. Bartynski, M. Bubak, G. Dyk, T. Gubala, and M. Kasztelnik. A Development andExecution Environment for Early Warning Systems for Natural Disasters. In Cluster, Cloud andGrid Computing (CCGrid) 2013, pages 575–582. IEEE, 2013.

[2] B. Balis, M. Kasztelnik, M. Bubak, T. Bartynski, T. Guba�la, P. Nowakowski, and J. Broekhuijsen.The UrbanFlood Common Information Space for Early Warning Systems. Procedia ComputerScience, 4:96–105, 2011.

[3] Bartosz Balis. Increasing Scientific Workflow Programming Productivity with HyperFlow. InProceedings of the 9th Workshop on Workflows in Support of Large-Scale Science, WORKS ’14,pages 59–69, Piscataway, NJ, USA, 2014. IEEE Press.

[4] P. Beckman, S. Nadella, N. Trebon, and I. Beschastnikh. Spruce: A system for supporting ur-gent high-performance computing. In Grid-Based Problem Solving Environments, pages 295–311.Springer, 2007.

[5] R. Brzoza-Woch, M. Konieczny, B. Kwolek, P. Nawrocki, T. Szydo, and K. Zieliski. Holisticapproach to urgent computing for flood decision support. Procedia Computer Science, 2015. Pro-ceedings of the International Conference on Computational Science, ICCS 2015.

[6] M. Chuchro, M. Lupa, A. Pieta, A. Piorkowski, and A. Lesniak. A Concept of Time WindowsLength Selection in Stream Databases in the Context of Sensor Networks Monitoring. In NewTrends in Database and Information Systems II, volume 312, pages 173–183. Springer, 2015.


2385

[7] W. Funika, K. Mazurek, and W. Kruczkowski. An agent-based hierarchical approach for executingbag-of-tasks in clouds. Computer Science, 15:1, 2014.

[8] Konstantin V Knyazkov, Denis A Nasonov, Timofey N Tchurov, and Alexander V Boukhanovsky.Interactive workflow-based infrastructure for urgent computing. Procedia Computer Science,18:2223–2232, 2013.

[9] J. Kosinski, P. Nawrocki, D. Radziszowski, K. Zielinski, S. Zielinski, G. Przybylski, and P. Wnek.SLA Monitoring and Management Framework for Telecommunication Services. In ICNS, pages170–175. IEEE Computer Society, 2008.

[10] S. V. Kovalchuk, P. A. Smirnov, S. V. Maryin, T. N. Tchurov, and V. A. Karbovskiy. Deadline-driven resource management within urgent computing cyberinfrastructure. Procedia ComputerScience, 18:2203–2212, 2013.

[11] V.V. Krzhizhanovskaya, G.S. Shirshov, N.B. Melnikova, R.G. Belleman, F.I. Rusadi, B.J.Broekhuijsen, B.P. Gouldby, J. Lhomme, B. Balis, M. Bubak, A.L. Pyayt, I.I. Mokhov, A.V.Ozhigin, B. Lang, and R.J. Meijer. Flood early warning system: design, implementation andcomputational modules. Procedia Computer Science, 4:106–115, 2011.

[12] Siew Hoon Leong, Anton Frank, and Dieter Kranzlmuller. Leveraging e-infrastructures for urgentcomputing. Procedia Computer Science, 18:2177–2186, 2013.

[13] M. Malawski, K. Figiela, and J. Nabrzyski. Cost minimization for computational applications onhybrid cloud infrastructures. Future Generation Computer Systems, 29(7):1786–1794, 2013.

[14] Maciej Malawski, Kamil Figiela, Marian Bubak, Ewa Deelman, and Jarek Nabrzyski. SchedulingMultilevel Deadline-Constrained Scientific Workflows on Clouds Based on Cost Optimization.Scientific Programming, 2015, 2015.

[15] W. J. Moscicki, G. Bania, M. Cwiklik, and A. Borecka. DC Resistivity Studies of Shallow Geologyin the Vicinity of Vistula River Flood Bank in Czernichow Village (near Krakow in Poland). StudiaGeotechnica et Mechanica, 36(1):63–70, 2014.

[16] A.-M. Oprescu, T. Kielmann, and H. Leahu. Budget estimation and control for bag-of-tasksscheduling in clouds. Parallel Processing Letters, 21(02):219–243, 2011.

[17] A Pieta, J Bala, M Dwornik, and K Krawiec. Stability of the levees in case of high level of the water.In 14th SGEM Geoconference On Informatics, Geoinformatics And Remote Sensing–ConferenceProceedings, volume 1, pages 809–815, 2014.

[18] Michael Schiffers, D Kranzlmuller, Andrea Clematis, Daniele D’Agostino, Antonella Galizia, Al-fonso Quarati, Antonio Parodi, Marina Morando, Nicola Rebora, et al. Towards a grid infrastruc-ture for hydro-meteorological research. Computer Science, 12:45–62, 2011.

[19] S Srinivasan. Risk management in the cloud and cloud outages. In Security, trust and regulatoryaspects of cloud computing in business environments, chapter 8. Hershey: IGI Global, 2014.

[20] R. Strijkers, W. Toorop, A. van Hoof, P. Grosso, A. Belloum, D. Vasuining, C. de Laat, andR. Meijer. Amos: Using the cloud for on-demand execution of e-science applications. In e-Science(e-Science), 2010 IEEE Sixth International Conference on, pages 331–338. IEEE, 2010.

[21] T. Szydlo, P. Nawrocki, R. Brzoza-Woch, and K. Zielinski. Power aware MOM for telemetry-oriented applications using GPRS-enabled embedded devices levee monitoring use case. In Proc.2014 Federated Conference on Computer Science and Information Systems, volume 2 of Annals ofComputer Science and Information Systems, pages 1059–1064. IEEE, 2014.

[22] Konstantin V Knyazkov, Sergey V Kovalchuk, Timofey N Tchurov, Sergey V Maryin, and Alexan-der V Boukhanovsky. Clavire: e-science infrastructure for data-driven computing. Journal ofComputational Science, 3(6):504–510, 2012.

[23] Urzad Statystyczny w Krakowie (Statistical Office of Krakow). Powodz w wojewodztwie malopol-skim w 2010 roku (Flood in the Malopolska province in 2010). http://krakow.stat.gov.pl/cps/rde/xbcr/krak/ASSETS_2011_skutki_powodzi_2010.pdf, 2011. In Polish. Accessed 7.01.2015.

[24] KK Yoshimoto, DJ Choi, RL Moore, Asmit Majumdar, and Eva Hocks. Implementations of UrgentComputing on Production HPC Systems. Procedia Computer Science, 9:1687–1693, 2012.


2386

execution management and efficient resource provisioning ...doi: 10.1016/j.procs.2015.05.412...

Documents