Precision-Aware application execution for Energy-optimization in HPC node system

0.61 MB·English

by  Radim Vavřík

Precision-Aware application execution for
Energy-optimization in HPC node system
Radim Vavrik Antoni Portero Stepan Kuchar
IT4InnovationsNational IT4InnovationsNational IT4InnovationsNational
SupercomputingCenter SupercomputingCenter SupercomputingCenter
VSB-TechnicalUniversityof VSB-TechnicalUniversityof VSB-TechnicalUniversityof
Ostrava Ostrava Ostrava
Ostrava,CzechRepublic Ostrava,CzechRepublic Ostrava,CzechRepublic
[email protected] [email protected] [email protected]
5
Martin Golasowski Simone Libutti Giuseppe Massari
1
IT4InnovationsNational PolitecnicodiMilano PolitecnicodiMilano
0
2 SupercomputingCenter DipartimentodiElettronica, DipartimentodiElettronica,
VSB-TechnicalUniversityof InformazioneeBiongegneria InformazioneeBiongegneria
n
Ostrava (DEIB) (DEIB)
a
Ostrava,CzechRepublic Milano,Italy Milano,Italy
J
[email protected] [email protected] [email protected]
6
2
ABSTRACT touseamulti-nodesolution. Theoptimaltrade-offbetween

] Power consumption is a critical consideration in high per- precisionvs. executiontimeiscomputedbytheRTRMwith
C
formancecomputingsystemsanditisbecomingthelimiting thetimeoverheadlessthan10%againstanativeexecution.
D
factortobuildandoperatePetascaleandExascalesystems.
. When studying the power consumption of existing systems CategoriesandSubjectDescriptors
s
runningHPCworkloads,wefindthatpower,energyandper-
c [HPCsystems,disastermanagement,reliabilitymod-
[ formancearecloselyrelatedwhichleadstothepossibilityto els, computing resource management, many-cores,
optimize energy consumption without sacrificing (much or parallel systems]: [PACS: 89.20.Ff, 47.50.Cd, 92.40.-t,
2 at all) the performance. In this paper, we propose a HPC
07.05.Bx, 07.05.Tp, 89.20.Ff]
v
system running with a GNU/Linux OS and a Real Time
7
ResourceManager(RTRM)thatisawareandmonitorsthe
5 Keywords
healthy of the platform. On the system, an application for
5 HPCsystems,disastermanagement,runtimeresourceman-
disaster management runs. The application can run with
4 agement
different QoS depending on the situation. We defined two
0
mainsituations. Normalexecution,whenthereisnoriskof
1. a disaster, even though we still have to run the system to 1. INTRODUCTION
0 look ahead in the near future if the situation changes sud- Applicationrequirements,power,andtechnologicalconstrain-
5 denly. Inthesecondscenario,thepossibilitiesforadisaster tsaredrivingthearchitecturalconvergenceoffutureproces-
1 are very high. Then the allocation of more resources for sors towards heterogeneous many-cores. This development
v: improving the precision and the human decision has to be is confronted with variability challenges, mainly the sus-
i takenintoaccount. Thepapershowsthatatdesigntime,it ceptibility to time-dependent variations in silicon devices.
X ispossibletodescribedifferentoptimalpointsthataregoing Increasing guard-bands to battle variations is not scalable,
r to be used at runtime by the RTRM with the application. duetotoolargeworst-casecostimpactfortechnologynodes
a This environment helps to the system that must run 24/7 around10nm. Thegoalofnextgenerationfirmwareistoen-
in saving energy with the trade-off of losing precision. The able next-generation embedded and high-performance het-
paper shows a model execution which can improve the pre- erogeneousmany-corestocost-effectivelyconfrontvariations
cisionofresultsby65%inaveragebyincreasingthenumber byprovidingDependable-Performance: correctfunctionality
of iterations from 103 to 104. This also produces one order and timing guarantees throughout the expected lifetime of
ofmagnitudelongerexecutiontimewhichleadstotheneed a platform under thermal, power, and energy constraints.
An optimal solution should employ a cross-layer approach.
Amiddle-wareimplementsacontrolenginethatsteerssoft-
ware/hardware knobs based on information from strategi-
cally dispersed monitors. This engine relies on technology
models to identify/exploit various types of platform slack
- performance, power/energy, thermal, lifetime, and struc-
tural (hardware) - to restore timing guarantees and ensure
the expected lifetime and time-dependent variations.
Dependable-Performance is critical for embedded applica-
tions to provide timing correctness; for high-performance 2. THESTATEOFTHEART
applications, it is paramount to ensure load balancing in OneofthegoalsforfutureHPCsystemsistodevelopauto-
parallelphasesandfastexecutionofsequentialphases. The matedframeworksthatusepowerandperformancetomake
lifetime requirement has ramifications on the manufactur- applications-aware energy optimizations during an execu-
ing process cost and the number of field-returns. The fu- tion. Techniqueslikedynamicvoltageandfrequencyscaling
turefirmwarenoveltymustrelyinseekingsynergiesintech- (DVFS)forreducingthespeed(clockfrequency)inexchange
niquesthathavebeenconsideredvirtuallyexclusivelyinthe for reduced power consumption are used [11]. And power
embedded or high-performance domains (worst-case guar- gating technique that allows reducing power consumption
anteed partly proactive techniques in embedded, and dy- by shutting off the current to blocks of the circuit that are
namic best-effort reactive techniques in high-performance). not in use. Different computations have different power re-
This possible future solutions will demonstrate the benefits quirements. ForcomputationswheretheCPUiswaitingfor
of merging concepts from these two domains by evaluating resourcesthefrequencycanbereducedtolowerpowerwith
key applications from both segments running on embedded minimal performance impact [2]. System scenarios classify
andhigh-performanceplatforms. Theintentofthispaperis systembehavioursthataresimilarfromamultidimensional
todescribethecharacteristicsandtheconstraintsofdisaster cost perspective, such as resource requirements, delay and
management(DM)applicationsforindustrialenvironments. energy consumption, in such a way that the system can be
Whendefiningtherequirementsandtheirevaluationproce- configured to exploit this cost. At design-time, these sce-
dure, a first analysis of the DM applications modules (HW narios are individually optimized. Mechanisms for predict-
platform,OSandRTengines,monitorsandknobs,reliabil- ing the current scenario at run-time and for switching be-
itymodels)isprovided. Powerconsumptionisacriticalcon- tweenscenariosarealsoderived. Thesearederivedfromthe
sideration in high performance computing systems and it is combinationofthebehaviouroftheapplicationandtheap-
becomingthelimitingfactortobuildandoperatePetascale plicationsmappingonthesystemplatform. Thesescenarios
and Exascale systems. When studying the power consump- areusedtoreducethesystemcostbyexploitinginformation
tion of existing systems running HPC workloads, we find about what can happen at run-time to make better design
power,energyandperformancearecloselyrelatedleadingto decisionsatdesign-time,andtoexploitthetime-varyingbe-
thepossibilitytooptimizeenergywithoutsacrificing(much haviour at run-time. While use-case scenarios classify the
or at all) performance. In this paper, we propose a HPC application’sbehaviourbasedonthedifferentwaysthesys-
system running with an RTRM that is aware and monitors tem can be used in its over-all context, system scenarios
the healthy of the platform. On the system, an application classify the behaviour based on the multi-dimensional cost
for disaster management is running. The application can trade-off during the implementation [9].
run with different QoS depending on the situation. We de-
fined two main situations. Normal execution when there is Togetpreciseresultsasfastaspossibleisthemainrequire-
noriskofdisaster,eventhoughwestillhavetorunthesys- ment for our testing application with the risk of disaster
temtolookaheadinthenearfutureifthesituationchanges scenario. Runing applications with such requirements is
suddenly. And the second scenario where the possibilities called urgent computing [12]. There are plenty of specifics
for a disaster are very high. Then the allocation of more for such kind of computing, e.g. need for infrequent, but
resources for improving the precision and human decision massivecomputationalperformance(HPC).Sincereserving
has to be taken into account. The paper shows that at de- computingcapacityonlyforurgentcasesiseconomicallyin-
sign time, it is possible to describe different optimal points efficient, the usage of cloud computing can be the solution.
that are going to be used at runtime by the RTRM with Ontheotherhand,therearelimitationsforcommunication-
the application. This environment helps to the system that intensive applications and guarantee the availability of the
must run 24/7 in saving energy with the trade-off of losing resources may be problematic.
precision.
2.1 RTRM:Run-TimeResourceManager
The paper shows a model execution which can improve the
The framework [1],[3] is the core of a highly modular and
precisionofresultsby65%inaveragebyincreasingthenum-
extensible run-time resource manager which provides sup-
ber of iterations from 103 to 104. This also produces one
port for an easy integration and management of multiple
orderofmagnitudelongerexecutiontimewhichleadstothe
applicationscompetingontheusageofone(ormore)shared
need to use a multi-node solution. The optimal trade-off
MIMDmany-corecomputationdevices. Theframeworkde-
between precision vs. execution time is computed by the
sign,whichexposesdifferentplugininterfaces,providessup-
RTRM with the time overhead less than 10% against a na-
port for pluggable policies for both resource scheduling and
tive execution.
the management of applications coordination and reconfig-
uration. Applications integrated with this framework get a
The paper is divided into nine sections, after the introduc-
suitableinstrumentationtosupportDesign-Space-Explorati-
tion the state-of the art follows. In section three, we intro-
on (DSE) techniques, which could be used to profile appli-
duce our driving disaster management example named the
cation behaviours to either optimize them at design time
uncertainty modelling of rainfall-runoff model. Section four
or support the identification of optimal QoS requirements
describes the Real Time Resource Manager (RTRM) that
goals as well as their run-time monitoring. Suitable plat-
monitors the sensors and knobs of the platform. Section
form abstraction layers, built on top of GNU/Linux kernel
five shows results of possible optimal computation points
interfaces, allow an easy porting of the framework to differ-
that are obtained at the design time. They are going to be
entplatformsanditsintegrationwithspecificexecutionen-
usedbytheRTRMatruntimeinnearfuture. Finally,there
vironments. WeuseacommonGNU/Linuxwithanefficient
is a section of conclusion and future work.
Run-TimeResourceManager(RTRM)[1],[3]thatexploitsa
Design-TimeExploration(DSE),whichperformsanoptimal and one of the quantiles selected from the Monte Carlo re-
quantization of the configuration space of run-time tunable sults. TheN-Scoefficientisoftenusedforestimatingpreci-
applications,identifyingasetofconfigurations. Theconfig- sionofgivenmodelbycomparingitsoutputwithmeasured
urationofarun-timetunableapplicationisdefinedbyaset databutcanbesuccessfullyusedforcomparisonofanytwo
of parameters. Some of them could impact the application model outputs.
behaviour (for example the uncertainty of Rainfall-Runoff
(RR) models can produce different accuracy for different
situations) while other have direct impact on the amount (cid:80)T (Qt −Qt )2
of required resources. For example, the amount of Monte E =1− t=1 0 m (1)
Carlo iterations in uncertainty modelling and the time be- (cid:80)Tt=1(Qt0−Qt0)2
tween batches of simulations lead to different requirements
for allocating system resources.
Where:
Q0 isthemeasuredflowinonetimestep.
3. DISASTERMANAGEMENTMODEL Qm isthesimulatedflowinonetimestep.
EistheNash-Sutcliffemodelefficiencycoefficient.
ThemodelunderstudyisamodularpartoftheFLOREON+
E=1meansthatthesimulationmatchesobserveddataperfectly.
project[4,5]. ThemainobjectiveoftheFLOREON+project E=0simulationmatchesmedianoftheobserveddata.
istocreateaplatformforintegrationandoperationofmon- E<0simulationislessprecisethanthemedianoftheobserved
itoring, modelling, prediction and decision support for dis- data.
astermanagement. ModularityoftheFLOREON+system,
which is developed for this science and research project, al-
Section 5 presents our study about the precision that the model
lows for simple integration of different thematic areas, re-
providesanditsrelationwiththeamountofcomputation. Aswe
gions and data. This system is meant to be used for de-
explainlater,atdesigntime,itispossibletosimulatesituations
cision support in operative and predictive disaster manage- with different trade-offs in terms of quality of services (QoS),
ment processes. The central thematic area of the project executiontimeandpower-consumption. ThentheRTRMchoose
is hydrologic modelling and prediction. The system focuses theoptimalworkingpointdependingontheapplicationscenarios.
onacquisitionandanalysisofrelevantdatainrealtimeand
application of prediction algorithms with this data. The 3.1 ApplicationScenarios
results are then used for decision support in disaster man- Application scenarios [10] describe different triggers and states
agementprocessesbyprovidingpredicteddischargevolumes of the application that influence the system responsiveness and
on measuring stations and prediction and visualization of operation (e.g. critical flooding level, critical state of patient’s
health,voice&data,etc.). Basedonthesescenarios,thesystem
floodlakesinthelandscape. Modelinthisstudy[7]isused
canbeindifferentstateswithdifferentservicelevelrequirements
forshorttermpredictionsofhydrologicalsituationongiven
thatcanalsobetranslatedtotheparametersoftheuncertainty
rivers. Rainfall-runoff (RR) model is a dynamic mathemat- modelling and its required resources. Shorter response time in
ical model which transforms rainfall to flow at the catch- criticalsituationscanforexamplebeacquiredbydecreasingthe
ment outlet. The main purpose of the model is to describe number of Monte Carlo iterations (also decreasing the precision
rainfall-runoff relations of a catchment area. Common in- of the results), or by allocating more computational resources
(maintaining the same precision level if the load is rebalanced
puts of the model are precipitations on given stations and
appropriately). Weidentifiedtwomainapplicationscenariosthat
spatial and physical properties of the river catchment area.
support the different workload of the system based on the flood
Commonoutputsaresurfacerunoffhydro-graphswhichde- emergencysituation.
pict relations between discharge (Q) and time (t). Catch-
mentareasforthemodelusedinthispaperforexperiments
3.1.1 StandardOperation
are parametrized by following values. Manning’s roughness
Inthisscenario,weatherisfavourableandthefloodwarninglevel
coefficent(N)whichapproximatesphysicalpropertiesofthe
isbelowthecriticalthreshold. Thecomputationcanberelaxed;
river basin and CN curve number which approximates ge- some level of errors and deviations can be allowed. The system
ological properties of the river catchment area. Rainfall- shouldonlyuseasmuchpowerasneededforstandardoperation;
runoff models are usually used for predicting behaviour of onebatchofRRsimulationswithuncertaintymodellingonlyhas
river discharge and water level and one of the inputs of to be finished before the next batch starts. The results do not
have to be available as soon as possible, so no excess use of re-
rainfall-runoffmodelsistheinformationaboutweathercon-
sourcesisneeded.
ditions in the near future. These data are provided by nu-
merical weather prediction models [7] and can be affected
3.1.2 EmergencyOperation
by certain inaccuracy. Such inaccuracy can be projected
Severaldaysofcontinuousrainraisethewaterinriversoravery
into output of the model by constructing confidence inter-
heavy rainfall on a small area creates new free-flowing streams.
valswhichprovideadditionalinformationaboutpossibleun-
Theseconditionsaresignalledbythedischargevolumeexceeding
certainty of the model output. The confidence intervals are thefloodemergencythresholdsorprecipitationamountexceeding
constructed by using iterative Monte Carlo method. Data the flash flood emergency thresholds. Much more accurate and
sets are sampled from the model input space and used as frequentcomputationsareneededinthisscenarioandresultshas
input for a large number of simulations. Quantiles selected to be provided as soon as possible even if excess resources have
tobeallocated.
fromthesimulationsoutputaremergedwiththeoriginalhy-
drograph. Precision of modelling of the model uncertainty
4. RTRMMETHODS:ABSTRACTEXECU-
can be positively affected by increasing the number of per-
formedsimulations. Theprecisionofthesimulationscanbe TIONMODEL
determined by estimating the Nash-Sutcliffe (N-S) model Thedevelopmentoftheruntimereconfigurableapplicationisfa-
efficiency coefficient between the original simulation output cilitated using a real time library (RTLIB). RTLIB provides the
Figure 2: Precision vs. number of iterations.
gionunderstudyreceivedfromtheweatherstations. New
dataisreceivedfromthereeachtenminutestoonehour.
Figure 1: Diagram of the RTRM methods for un-
• Threadsforuncertainty: Thenumberof(OpenMP)threads
certainty Application. equivalenttothenumberofMonteCarloexecutionsofthe
rainfall-runoff model that runs in parallel to simulate the
uncertainty.
Abstract Execution Model (AEM). The AEM encapsulates the
execution flow of the application in a way that let the real time
Threads for rainfall-runoff model: The number of (OpenMP)
resource manager RTRM [1] to govern its life-cycle. Generally,
threadsthatparallelizethemainloopsoftherainfall-runoffmodel
wecanstructureanapplicationbysplittingitinmoreExecution
itself.
Contexts(EXC).FromtheRTRMpointofviewanEXCisseen
as an schedulable task. Having more EXC in the same applica-
tioncancomefromtheneedofschedulepartsoftheapplication MethodonConfigure
withdifferentprioritiesandresourceusages. Thisismadepossi-
CalledwhentheRTRMisassignedforthefirsttime,ortheAp-
bleusingtherecipethatisassociatedtotheExecutionContexts
plication Working Mode (AWM) has changed, i.e., the set of
[8].
resources allocated for the application/EXC. The code of this
method is related to whatever is required to reconfigure the ap-
Fromtheapplicationside,themandatoryrequirementtoexploit
plication(numberofthreads,applicationparameters,datastruc-
the RTRM is to provide an implementation that allows it to be
tures,etc.) toproperlyrunaccordingtoresourcesassignedthrough
run-time reconfigurable. From the RTRM side, the configura-
theAWM.
tionofanapplicationischaracterizedintermsofsetofresource
requirements, and is called Application Working Mode (AWM).
TheapplicationmustdeclaretotheRTRMthelistofitsAppli- MethodonRun
cationWorkingModes. Suchlistmustbeprovidedthroughafile
This is the entry point of the task. In our case the uncertainty
called, in jargon, Recipe which basically is an XML file. Fig. 1
simulation. Thereisacalltoourwrapped’main’methodcalled
summarizes the Abstract Execution Model, where the Applica-
’floreon-model’. It has a structure ( named val) as an input,
tion,onConfigure,WorkloadTerminated? andQoS-OK?boxes
whichprovidesinputparametersthatcanvary(modify)theper-
are related to the member methods of the derived class. As we
formanceofthemodel.
can see, the AEM configures a sort of state diagram. Thus the
methods must implement what our Execution Context must do
wheneverreachthecorrespondingexecutionstate.
(Result,Quality)=f[floreon-model(val)]; (2)
MethodonSetup Where
Quality=f(E,t); (3)
InitializationoftheframeworkisperformedintheonSetup func-
tion. In order to allow the RTRM to perform a correct initial- E-Precision(Nash-Sutcliffe model efficiency coefficient), t-time
izationofthestructuressupportingthecollectionoftheruntime
statistics, this method performs the initialization of the struc-
turenamed”val”. Ithasalltheinputvaluesthatcanmodifythe MethodonMonitor
uncertaintysimulation. Thedescriptionofthevaluesis: After a computational run, the application may check whether
the level of QoS/performance/accuracy is acceptable or not. In
the second case, some action could be taken. This method is
• CN, N range and precision: Probabilitydistributionfunc-
in charge to monitor the results received from the uncertainty
tionanditsparametersusedforgeneratingsamplesofthe
model. ItdecidesifthecurrentAWMisenoughtocomputenext
CNnumberandManning’srougnesscoefficientvalues.
uncertaintyiterationortheAWMandinputdata(valstruct)has
• Input Rain Data: The data about rainfall on the land re- tobemodifiedtocorrespondthenewsituationatruntime.
MethodonRelease
Optional,butrecommendedmemberfunction. Thisenvironment
shouldeventuallycontainclean-upoftheframeworkandapplica-
tions. Thisistheexitmethodtoexecutewhenthecomputation
finalizes. It is the end of the program and the use must be an
user decision. The system eventually does not have to arrive to
thispointsincethesystemrunning24/7isexpected.
5. RESULTS
Therearetwoplatformsthatservefordevelopment,testingand
integration of the example application exposed above. First is a
48coreSMPmachineHPProLiantDL785G6[6]. AGNU/Linux
version 3.13.0-36-generic with a gcc version 4.8.2 is installed on
thismachine. Itconsistsof8socketswithSix-CoreAMDOpteron(tm)
Processor 8425 HE 64bits 2.1@GHz clock frequency. Each core
hasaL1cachememorysizeof64KBdataand64KBforinstruc-
tionsand512KBL2cachememoryandaTLBof1024KB.Each
CPUsockethasanembeddedmemory5118KBL3. Themachine
isequippedwith256GBDDR2533MHzwithECCmemory. The
RTRMsystemusesresourcesoftheplatform,8socketseachwith
six cores in total. One of the cores of each socket, six in total,
named hosts, are allocated to monitor the system and the sen-
sorsoftheplatform(i.e. power,temperatureetc.) bytheRTRM
system and to run all non-managed applications and processes.
Therefore,therearejust42realcoresforexecutingmanagedap-
plicationswiththeRTRMsupport.
Second platform is HPC cluster Anselm, which consists of 209
compute nodes, totaling 3344 compute cores. Each node is a
Figure3: Uncertaintyexecutednatively,timeversus
powerful x86-64 computer, equipped with 16 cores and at least
energy consumption, on SMP machine.
64GBRAM.Nodesareinterconnectedbyfullynon-blockingfat-
tree Infiniband network and equipped with Intel Sandy Bridge
processors. TheclusterrunsbullxGNU/Linuxoperatingsystem,
whichiscompatiblewiththeRedHatGNU/Linuxfamily. There
isnoRTRMsupportmeantime,butitispartoffuturework.
After integration of the uncertainty modelling module into the
RTRM on SMP machine, we first computed how the coefficient
to compute the precision of the simulations decrease with the
numberofiterations. Thiscoefficientiscomputedandcumulated
ineachiterationandheadstothezeroatinfinity. Lowervalueof
thecoefficientmeanshigherprecision. Theseprecisionsforgiven
numbersofiterationsasFig. 2depictsareresourcesindependent
andwillbeusedtoidentifyoptimalworkingmodes.
Fig. 3 presents energy consumption of native, which means un-
managed by RTRM, executions of uncertainty simulations with
agalaxyofpoints. Valuesareonlyestimatedbecauseofabsence
ofHWsupportforpowerconsumptionmeasurementoftheSMP
machine. Thesepointsrepresentsimulationswithdifferentpreci-
sion-numberofiterations(e.g. 100lowprecision,1000medium
precisionor10000highprecision),differentCPUfrequenciesand
variousnumberofcoreswhichresultsindifferentexecutiontimes
and appropriate consumed energy. This leads to trade-offs for
runningthesystemwithvariousamountofresources(cores,fre-
quencies). Fig. 4 shows similar results but instead of native
executionthesimulationsrunwiththeRTRM.
Dynamicpart(onlyusedcores)ofestimatedpowerconsumption
wascomputedusingoriginalformula:
n
Pn=(Pmax−Pidle)100 +Pidle (4)
Where: n is a system load, Pn is a power consumption at this
load, Pmax is a maximum power consumption, Pidle is a power
consumptioninanidle. Figure 4: Uncertainty executed with RTRM, time
versus energy consumption, on SMP machine.
EstimatedenergyconsumptionE isthencomputedfromthefor-
mula:
E=Pt (5)
Where: P is estimated power consumption and t is time of exe-
cution.
Figure 5: Comparison of RTRM time overheads for Figure6: Uncertaintyexecutednatively,timeversus
different configurations. energy consumption, on HPC cluster.
8. ADDITIONALAUTHORS
We have observed that the overhead of RTRM in time is below
10% in average, which is shown by Fig. 5. We have to take Additional authors: William Fornaciari (Politecnico di Milano,
intoaccountthatfromeachsocket(sixx86-64cores)oftheplat- email: [email protected]) and Vit Vondrak (IT4
form,onecoreisusedtomonitorthehealthyofthesystem. So, InnovationsNationalSupercomputingCenter,email:
from48possiblecoresonly42areenabledtorunouruncertainty [email protected]).
module. From these points is possible to extract which are the
optimal-paretopointsthataregoingtobeinsertedintheRTMR 9. REFERENCES
recipe. Forcomparison,powerconsumptionofnativeexecutions
[1] BellasiP.etal.: Artrmproposalformulti/many-core
of uncertainty simulations on HPC cluster (multi-node solution
platformsandreconfigurableapplications.InReCoSoC,
usingMPI+OpenMP,10000iterations)areshownonFig. 6. Val-
2012.
uesweremeasuredusingtheLikwid(likwid-powermeter)toolfor
different numbers of nodes. The graph shows steeply increasing [2] TiwariA.,etal.: GreenQueue: CustomizedLarge-scale
power consumption with no significant time improvement using ClockFrequencyScaling.CGC2012
morethansixteen16-corescomputingnodes. [3] Harpaharnessingperformancevariabilityfp7project,
http://www.harpa-project.eu,2013.
[4] MartinovicJ.,etal.: MultipleScenariosComputingInThe
6. CONCLUSIONANDFUTUREWORK
FloodPredictionSystemFLOREON.ECMS2010: 182-188
The paper presents an environment that helps to an x86-64 ar- [5] TheuerM.,etal.,Efficientmethodsofautomatic
chitecture platform to run 24/7 a model saving energy with the calibrationforrainfall-runoffmodellingintheFloreon+
trade-off of losing precision with the different resources config- system,InternationalJournalonnon-standardcomputing
urations. The paper shows the uncertainty modelling executed andartificialintelligence(Accepted),2014
with different steps of precision given by various numbers of it- [6] HPProLiantDL785Generation6(G6)Quick
erations leading in appropriate execution times. The decision Specificationshttp://www8.hp.com(docname=c04284472)
of precision/executed time/consumed energy is taken by a Run- (October2014)
TimeResourceManagermonitoringtheworkloadoftheplatform
[7] GolasowskiM.etal.: Uncertaintymodellingin
andcomputingthebest-tradeoffwiththetimeoverheadlessthan
Rainfall-RunoffsimulationsbasedonparallelMonteCarlo
10%inagainstanativeexecution. At designtime, itispossible
method,InternationalJournalonnon-standardcomputing
to describe the optimal points that are going to be used at run-
andartificialintelligence(accepted)2014.
time. Theenvironmentalsopermitstoexecutethemodule24/7
[8] BOSP:TheBarbequeOpenSourceProject
in nodes of a cluster and/or with accelerators. This enables to
http://bosp.dei.polimi.it/doku.php?id=how-to
usemoreresourcesandhigherprecisionforthefuture.
[9] GheorgitaS.V.,etal.: Systemscenariobaseddesignof
dynamicembeddedsystems.TransactionsonDesign
7. ACKNOWLEDGMENTS AutomationofElectronicSystems(ToDAES),Volume14
ThisarticlewassupportedbyOperationalProgrammeEducation Issue1,ArticleNo.3,2009.
forCompetitivenessandco-financedbytheEuropeanSocialFund [10] PorteroA.,etal.: SystemandApplicationScenariosfor
within the framework of the project New creative teams in pri- DisasterManagementProcesses,theRainfall-RunoffModel
oritiesofscientificresearch,reg. no. CZ.1.07/2.3.00/30.0055,by CaseStudy.CISIM2014: 315-326
theEuropeanRegionalDevelopmentFundintheIT4Innovations [11] AnagnostopoulosI.,etal.Power-awaredynamicmemory
Centre of Excellence project (CZ.1.05/1.1.00/02.0070), by the managementonmany-coreplatformsutilizingDVFS.ACM
project Large infrastructures for research, development and in- Trans.Embed.Comput.Syst.13,1s,Article40(Dec.2013)
novation of Ministry of Education, Youth and Sports of Czech [12] GalanteG.etal.,ASurveyonCloudComputingElasticity,
Republic with reg. no. LM2011033, and by 7th EU framework IEEEFifthInternationalConferenceonUtilityandCloud
programme project no. FP7-612069 HARPA - Harnessing Per- Computing,263-270,2012
formanceVariability.