Towards exascale real-time RFI mitigation

2.5 MB·

by  Rob V. van Nieuwpoort

TOWARDSEXASCALEREAL-TIMERFIMITIGATION
RobV.vanNieuwpoort
NetherlandseSciencecenterandUniversityofAmsterdam,TheNetherlands
[email protected],[email protected]
7
1 ABSTRACT frequency resolutions, before any integration is done, lead-
0 ing to more accurate results. Most importantly, we can re-
We describe the design and implementation of an extremely
2 move RFI before beam forming, which combines data from
scalable real-time RFI mitigation method, based on the of-
n all receivers. With beam forming, not only the signals, but
flineAOFlagger. Allalgorithmsscalelinearlyinthenumber
a alsotheRFIthatispresentinthedatastreamsfromthesepa-
J ofsamples. Wedescribehowweimplementedtheflaggerin
7 theLOFARreal-timepipeline,onbothCPUsandGPUs. Ad- ratereceiversiscombined,effectivelytakingtheunionofall
RFI.Thus, theRFIfromanyreceiverwillpolluteallbeams.
2 ditionally,weintroduceanovelsimplehistory-basedflagger
Therefore, it is essential to also perform real-time RFI miti-
thathelpsreducetheimpactofoursmallwindowonthedata.
] gationbeforethebeamformer,eventhoughthedataratescan
M By examining an observation of a known pulsar, we
be very high at this point. This is particularly important for
demonstratethatourflaggercanachievemuchhigherquality
I pulsarsurveys,forinstance.
thanasimplethresholder,evenwhenrunninginrealtime,on
h. The algorithms we use are based on earlier work by Of-
adistributedsystem. Theflaggerworksonvisibilitydata,but
p fringa and others [1, 2, 3]. Although our techniques are
alsoonrawvoltages,andbeamformeddata. Thealgorithms
- generic,wedescribehowweimplementedreal-timeRFImit-
o arescale-invariant,andworkonmicrosecondtosecondtime
igationforoneoftheSKApathfinders: TheLowFrequency
r scales. We are currently implementing a prototype for the
t Array(LOFAR)[4]. ThemodifiedRFImitigationalgorithms
s timedomainpipelineoftheSKAcentralsignalprocessor.
a we introduce here are extremely fast, and the computational
[
IndexTerms— RFI,real-time,LOFAR,SKA,CSP,TDT requirementsscalelinearlyinthenumberofsamplesandfre-

1 quency channels. We evaluate the quality of the algorithms
v with LOFAR pulsar observations. Using the signal-to-noise
1. INTRODUCTION
7
ratios of the folded pulse profiles, we can qualitatively and
9
1 Radio Frequency Interference (RFI) mitigation is extremely quantitatively compare the impact of different real-time RFI
8 important to take advantage of the vastly improved band- mitigationalgorithms.
0 width, sensitivity, and field-of-view of exascale telescopes. InadditiontoCPUversions,wehavedevelopedaproto-
1. Forcurrentinstruments, RFImitigationistypicallydoneof- typeforGraphicalProcessingUnits(GPUs). Wepresentthe
0 fline, and in some cases (partially) manually. At the same very promising performance results of performing real-time
7 time,itisclearthatduetothehighbandwidthrequirements, RFI mitigation on GPUs. Finally, we are now working on
1 RFI mitigation will have to be done automatically, and in incorporating our CPU and GPU codes in the SKA Central
:
v real-time,forexascaleinstruments. Signal Processor (CSP), in the context of the Time Domain
Xi In general, real-time RFI mitigation will be less precise Team(TDT)pulsarandtransientpipeline.
thanofflineapproaches. Duetomemoryconstraints,thereis
r
a much less data to work with, typically only in the order of 2. LOF:THELOFARONLINEFLAGGER
one second or less, as opposed to the entire observation. In
addition,duetomemorylimitationsandthefactthatprocess- For the offline RFI mitigation for the LOFAR telescope, we
ing is typically done in a distributed system, we can record usetheAOFlagger[3]. Thisflaggerisrelativelyefficient,and
onlylimitedstatisticsofthepast. Moreover,wewilltypically is fast enough to be applied in modern high-resolution ob-
have only few frequency channels locally available at each servatories. The AOFlagger operates on visibility data, and
computenode. Finally,theamountofprocessingthatcanbe currently is used only in the imaging mode. Although the
spentonRFImitigationisextremelylimitedduetocomput- AOFlaggerprovidesaselectionofmanydifferentalgorithms,
ing constraints and a limited power budget. Many existing for online use, we ported the most important and successful
algorithmsarethereforefartooexpensiveandnotapplicable. onestotheLOFARreal-timecentralprocessingsystem,that
Nevertheless, there are many potential benefits as well, originally ran on an IBM Blue Gene/p supercomputer. We
which include the possibility of working on higher time and alsoportedthealgorithmstoNVIDIAGPUsusingCUDA,as
2.2. Thescale-invariantrankoperator
Fig. 1. Left graphs shows the default factor for different
I
iterations;rightgraphshowstheoperationofSumThreshold.
Theslidingwindowdoubleseveryiteration.
Fig.2.ResultofrunningtheSIRoperator.Purplesamplesare
flagged by SumThreshold; yellow samples are subsequently
flaggedbytheSIRoperator. ImagecourtesyAndreOffringa.
described in Section 6.1. We call our implementation of the
real-timeflaggerLOF,shortforLOFAROnlineFlagger.
The scale-invariant rank operator [2] makes SumThresh-
oldmorerobust, byextendingtherangesofflaggedsamples
byapercentageofthesizeoftheflaggedrange.Atypicalper-
2.1. TheSumThresholdalgorithm centage we used is 20%. This means that all ranges of con-
secutiveflaggedsamplesareextendedbyflagging20%more
ThemostimportantalgorithmweuseisSumThreshold[1]. It samples, before and after the original range. The algorithm
performsthresholdingwithanexponentiallyincreasingwin- canberuninboththetimedirectionandthefrequencydirec-
dowsize,andanincreasinglysharperthreshold. Thisway,it tion. TheSIRoperatorhelpstoremoveRFIthatslowlyrises
candetectRFIatdifferentscales. Aswewilldemonstratein anddecreasesinstrength,thatmaybeotherwiseundetected.
Section5,thisindeedworkswellinpractice,andeffectively TheresultoftheSIRoperatorisshowninFigure2.
removes RFI from microsecond to multiple second scales. WeuseaversionoftheSIRoperatorimplementationthat
With SumThreshold, we define the threshold for the current hasalinearcomputationalcomplexityinthenumberofsam-
windowasfollows: ples(theoriginalimplementationintheAOFlaggerhadworse
thresholdI =median+stddev∗factorI ∗sensitivity computational complexity). The linear version of the algo-
Wherethefactoris: factor = startThreshold∗p2log(I) rithmisdescribedin[2].SinceSIRoperatesontheflagmasks
I I
only,andnotontheactualdataitself,itisextremelyefficient.
Typical values are p = 1.5, and sensitivity = 1.0. All mea-
surements use the defaults, since we empirically found that
3. INTEGRATIONINREAL-TIMEPIPELINES
theyprovideoptimalresults. Notuningisrequired. Figure1
showsthethresholdfordifferentiterations.
Figure1showstheoperationoftheSumThresholdalgo-
rithm. Withthedefaultvalues, thealgorithmbeginsthefirst
iteration as a simple thresholder, flagging all single samples
thataremorethan6sigmaawayfromthemedian.Thesecond
iterationdoublesthewindowsizetotwosamples,butlowers
the threshold to 4.5 sigma. We typically run 7-10 iterations,
dependingontheresolutionandsizeoftheinput.
SumThresholdcanberuninaone-dimensionalmode,as
shown in Figure 1. This can then be done in both time and
frequencydirections. Alternatively,thealgorithmcanoperate
on2Ddatainonepassaswell(notshowninFigure1).Aswe
willexplaininSection4,forreal-timeuse,wedidimplement
a2Dcode,butwewillmostlyusetheonedimensionalversion
of the algorithm for performance reasons. It is important to
notethatthecomputationalcomplexityoftheSumThreshold
algorithm itself is linear in the number of samples. There- Fig.3. TheLOFARreal-timepipelines. Inputdataratesvary
fore,thealgorithmissuitableforreal-timeuse,whereweare between400gbit/sand1.6tbit/sforthepre-correlationflag-
limitedbythenumberofcomputecycleswecanspend. ger,dependingonthemode(4,8or16bitsamples).
Figure3showsahigh-leveloverviewoftheLOFARreal-
timecentralsignalprocessingpipeline. Theentirepipelineis
implementedinsoftware,andisdescribedindetailin[4],in-
cludingadetailedperformanceanalysis. TheLOFARonline
flaggercomponentsareplacedinfourdifferentplacesinthe
pipeline. Dependingontheconfigurationandtheobservation
type,oneormoredifferentflaggersareused.
Data arrives in the form of raw voltages at the top left
of the figure. Next, a number of steps are executed inde-
pendently of whether we are in imaging or beam forming
mode. The most important step is a polyphase filter bank
that splits the broad input subbands in narrower frequency
Fig.4. Integrationinthetimedirectiontoperformflaggingin
channels. Typical channel bandwidths are between 0.8 kHz
thefrequencydirectionwithhighINR.
and12kHz,withsampleratesbetween82microsecondsand
1.3 milliseconds. When we need extremely high time res-
olution (e.g, for millisecond pulsars or for the cosmic ray Thisapproachhastwobenefits: itimprovestheINR,whileat
pipeline),webypassthepolyphasefilterbankaltogether. For thesametimereducingthecomputationalcosts. Wefirstflag
thiscasewecreatedaspecialhightimeresolutionflagger. inthefrequencydirection. Thisremovesstrongnarrow-band
Next,thebandpassofthefirstpolyphasefilterbankthat RFIthatwouldotherwisedecreasethequalityofthestatistics
runs inside the stations on FPGAs is corrected. After this used to compute the thresholds. We found that this is fre-
bandpass correction, we inserted our pre-correlation flagger quently present in the LOFAR RFI environment. This also
that works on the channelized raw voltage data. It is im- isthereasonwecreatethenarrowapproximately1kHzchan-
portant to do this after the band pass correction, as this en- nels. Wehaveanalternativemethodthatimplements2Dflag-
suresthatthesensitivityisequalacrossallchannels. Thepre- ging,whilepartiallyintegratinginoneorbothdimensionsto
correlationversionisthemostimportantreal-timeflagger,es- improveINRandtoreducecomputecosts.
peciallyforthebeamformingmodes. Thebeamformerdoes All flaggers have linear computational complexity in the
aweightedadditionofthedatastreamsfromthedifferentsta- numberofsamples, regardlessoftheirplaceinthereal-time
tions,essentiallytakingtheunionofallRFIfromallstations. pipeline. Forinstance,thepre-correlationflaggerhasacom-
IfRFIispresentatastation,thiswillpollutealloutputbeams. plexityofO(nrStations*nrPolarizations*nrChannels*nr-
EspeciallyforuncorrelatedRFIandlongbaselines,thisissub Times),thepost-correlationflaggerisO(nrBaselines*nrPo-
optimal. larizations*nrChannels),andthepostbeamformingflagger
In addition, we can use our real-time flagger after the isO(nrBeams*nrChannels*nrTimes).
correlator, for real-time image-based transient detection, for
example. The drawback of performing RFI mitigation after
4.1. Usinghistoricalinformation
the correlator is that, depending on the number of baselines
andtheintegrationtime,dataratescanbehigherthanbefore
Oneofthemostdifficultproblemswithreal-timeRFImitiga-
thecorrelator. Finally, wehaveapostbeamformingflagger
tionistheverylimitedwindowontheobservation. Typically,
thatcanpotentiallybenefitfromabetterinterference-to-noise
wecanonlykeeponesecondofdataorless(e.g.,atenthofa
ratio (INR). Moreover, depending on the number of output
second)inmemory. Similarly, wecanonlykeepaverylim-
beams,thistypicallyrunsonlowerdatarates.
itednumberoffrequencychannelsinmemory. Thisisdueto
memoryconstraints,andpartiallybecauseweprocessthedata
4. CHANGESFORREAL-TIMEUSE on a distributed system. We create parallelism by perform-
ingdomaindecomposition. Thistypicallymeansthatdiffer-
To makesure the algorithmsused in theAOFlagger workin ent frequency subbands are processed on different compute
a real-time context, we had to make several changes. First, nodes. Finally, in some cases, processing a second of input
dependingontheinputdatatypeoftheflagger,wemayhave datatakeslongerthanasecond. Tostillmeetthereal-timere-
tocomputeamplitudesfirst. Thisisthecaseforthehigh-time quirements,subsequentsecondsareprocessedpartiallyover-
resolution flagger, and for the pre-correlation flagger. The lappingintime,bydifferentcomputenodes. Allthesefactors
post-correlationflaggerrunsonvisibilitydatadirectly,while severelylimitoursituationalawareness.
thepost-beamformingflaggerrunsonStokesIdata. Our solution for this problem is the introduction of a
In pre-correlation mode, we typically integrate the data. novelhistoryflaggerthatperformssimplethresholdingofthe
Wefounditparticularlyusefultointegratethetimedirection currentdatachunk,basedonstatisticsofpastchunks. Pseudo
fully for frequency flagging, and to integrate the frequency codeforthishistoryflaggerisshownbelow.
directionfullyfortimedomainflagging. Figure4showsthis.
Fig.5. Thehistorybuffer.
Fig.6. ThresholdingvsLOF,onrawvoltages. Left: original
// For all channels, we do the following:
// Keep a history buffer (sliding window) of data;right: differencethreshold LOF,i.e. RFIresidualsthat
// means of unflagged samples of the past seconds areflaggedbyLOF,butnotbysimplethresholding.
currentValue = meanOfUnflaggedSamples()
historyMean = meanOfMeans() = 15.5 MByte per second. If we want to keep 5 minutes of
historyStddev = stddevOfMeans() history in the buffer, we already need 300 samples, leading
threshold= historyMean + sensitivity * to a storage requirement of 4.5 GBytes. After the correla-
historyStddev
tor, requirements are even higher: baselines ∗ subbands ∗
if(currentValue < threshold) { channels ∗ 32bits = 2080 ∗ 248 ∗ 256 ∗ 4 = 504 MByte
addToHistory(station, subband, currentValue) per second, which, even if we want to store only 5 minutes,
} else {
already leads to 148 Gbytes of statistics data. Therefore, in
addToHistory(station, subband, threshold)
practice, even keeping these very limited statistics of only
flagThisIntegrationTime()
} a few minutes of the past already is extremely difficult, and
AsshowninFigure5,ourflaggerusesahistorybufferthat down-samplingmaybeneeded.
storesthemeansoftheunflaggedsamplesoftheprevioussec-
ondsofdata. Thisbufferessentiallyisaslidingwindowover
5. EVALUATION
the data. We use the buffer to compute the mean and stan-
darddeviationofthepreviousseconds,togiveusaframeof
In this section, we will present a qualitative and quantitative
referencefortheoverallsignalstrengthofthecurrentsecond.
evaluationofthereal-timeflagger,usingaLOFARpulsarob-
This is especially important, since a strong broadband RFI
servation. Weusethepulsarpipeline,becauseitallowsfora
eventthatlastslongerthanourintegrationtimecanotherwise
quantitative comparison: we perform dedispersion and fold-
notbedetected.
ing to create a folded pulse profile. Next, we compute the
SNRofthepulseprofileasameasureofquality. BetterRFI
4.2. Statisticsinareal-timepipeline
mitigationdirectlyleadstoahigherSNR.
Thequalityofthestatisticsthatwecomputeandkeepisim-
portant, especially since the window on the data is so small.
5.1. Apulsarobservation
Weuseonlyverybasicstatistics,suchas(winsorized)means,
medians, standard deviations, and the MAD (Median Ab- We performed an observation of pulsar B1919+21, which
solute Deviation). To compute the medians, which can be has a period of 1.3373 s, a pulse width 0.04 s, and a disper-
expensive, we use an efficient O(N) implementation (note sion measure (DM) of 12.455. We observed at 138.0145.2
thatsortingthedataalreadyisO(NlogN)atbest. Morecom- MHz(32subbands)with5stations: CS005, CS006, RS205,
plications result from using a distributed platform: statistics RS406,andUK608. Wedeliberatelychoseanumberofcore
are often in the wrong place, at the wrong time. More- stations,whereRFIisexpectedtobecorrelated,aremotesta-
over, we cannot compute running statistics, since a second tionintheNetherlands,wheretheRFIenvironmentisknown
of data takes more than a second to compute. Finally, our tobeparticularlybad,andaninternationalstationintheUK,
real-timepipelineusescomplexcommunicationpatternsdue toguaranteewealsohaveuncorrelatedRFI.
to scheduling, and asynchronous communication for better WeusedaspecialLOFARmodethatallowedustostore
performance. Together,thismeansthatevencomputingbasic the raw UDP network packets, before the data even enters
statisticsisquitecomplexinpractice. the correlator. This allows us to replay the entire real-time
Animportantconsiderationforthehistoryflaggerispre- pipeline in an offline mode, enabling comparisons between
sented by the space requirements of the statistics of the past differentflaggingalgorithms,parametersettingsandevenob-
wewanttokeepinthehistorybuffer.LetususeLOFARnum- servation modes (e.g., imaging or beam forming). For the
bersasanexample. Forthepre-correlationflagger, weneed beam forming mode that we use for the pulsar pipeline, we
stations∗subbands∗channels∗32bits=64∗248∗256∗4 splitthefrequencysubbandsinto16channels(12KHz/82s).
older, the pulse is visible, but there also are false positives,
caused by strong RFI events that were not removed. With
LOF,thereisoneclearpeakwiththerightshape,withagood
SNR. This can be seen better in the right side of Figure 9,
which shows the same data, but zoomed in, with the non-
flaggedlineremoved.
Fig. 7. Beam formed data (Stokes I), top panel without RFI
mitigation,bottompanelwithLOF.
Fig.10. QualityofLOFforpulsarB1919+21. LOFhassig-
nificantlyhigherqualitythansimplethresholding. Thequal-
ityofLOFisclosetoofflinePrestorfifind.
Figure10showsthatLOFflags2.9%ofthedataintheob-
servation,significantlymorethanthesimplethresholderthat
onlyflags1.7%. Inthebottomleftgraph,weshowthatLOF
flagsawayabout15%ofthetotalsignalpower,onlyslightly
more than the simple thresholder. The bottom right graph
Fig. 8. Closeup of beam formed data (Stokes I), top panel
shows the SNR of the folded pulse profile. LOF is signifi-
withoutRFImitigation,bottompanelwithLOF.
cantlybetterthanthesimplethresholder, andalmostashigh
asperformingofflineRFImitigationwithPresto’srfifind[5].
6. TOWARDSEXASCALE
TheLOFARonlineprocessorneedshundredsofteraflopsof
computationalpower. FutureinstrumentssuchastheSquare
KilometreArray(SKA)willbemuchmoresensitive,andwill
Fig.9. Foldedpulseprofileswithandwithoutflagging.
requireordersofmagnitudemoreprocessing[6]. Therefore,
wewerecarefultomakesurethatallalgorithmsusedinour
Figure 6 shows two waterfall plots. The left side is the real-timeflaggerhavealinearcomputationalcomplexity, al-
original observation with RFI present; on the right the im- lowingexcellentscaling. Inaddition,weinvestigatedtheuse
provement of our LOF compared to a simple thresholding of modern processing architectures, such as GPUs. For LO-
scheme, where we manually determined the optimal thresh- FAR,wealreadyswitchedfromanIBMBlueGene/psystem
old. ThereclearlyisalotofresidualRFIthatisnotremoved toaGPUcluster.ForthecentralsignalprocessoroftheSKA,
bythethresholder,thatisremovedbyLOF. acombinationofFPGAsandGPUsarelikely. Inthissection,
wepresentaprototypeGPUversionoftheLOF.
Figure7showsStokesIdataofanoutputbeam;Figure8
shows the same, but zoomed in. The top panels are without
RFImitigation,thebottompanelswithLOF.AlmostallRFI
6.1. GPUimplementation
isremoved.Thedataisnotde-dispersed,butthepulsarsignal
issostrongthatitisclearlyvisibleinthedata. Notethatthe With Linus Schoemaker [7], we worked on the GPU port of
pulsesarenotflaggedawaybyourmitigationalgorithms. LOF.Duetothelimitedspace,wewillonlydescribethemost
InFigure9(leftside),weshowthefoldedpulseprofiles, important differences with the normal parallel CPU version
without RFI mitigation, with a simple thresholder, and with here. The normal LOF exploits parallelism by scheduling
LOF.WithoutRFImitigation,thepulsarsignaliscompletely different subbands to different compute nodes. Inside the
belowthenoisefloor,andcannotbedetected.Withthethresh- computenodesweuseC++andOpenMPformulti-threading,
7. CONCLUSIONSANDFUTUREWORK
We have demonstrated that our online flagger can achieve
much higher quality than simple thresholding, in real time,
even on a distributed system. The SumThreshold algorithm
was originally used mostly on visibility data. In this paper,
we demonstrated that the algorithm also works well on raw
voltages, pre-correlation data, and post-beam forming data.
Therefore,wehaveonerobustalgorithmforextremelydiffer-
entscales,frommicrosecondstomultipleseconds. Thealgo-
rithmsarescalableandhavelinearcomputationalcomplexity,
adding little overhead to existing pipelines. One complica-
tion is that we have an extremely limited view on our data,
and therefore need a history flagger. Due to the high data
rates,wehavetobeflexibleinstoragerequirements,evenfor
statistics.
WearecurrentlyworkingoncommissioningoftheGPU
code for LOFAR. Moreover, we are constructing a perfor-
Fig.11. ScalabilityoftheGPUimplementationofLOF.
mance model that can extrapolate Scalability towards SKA
sizes. This model includes power dissipation as well, since
thiswillbeanimportantbottleneckfortheSKA.Weplanto
andasynchronousMPImessagesforinter-nodecommunica-
usetheDomeExaBoundstoolforthisanalysis[8]. Allcode
tion. To avoid synchronization and parallelization overhead,
usedinthispaperisavailableasopensource:
we make the parallelism as coarse grained as possible. This
https://github.com/NLeSC/eAstroViz.
meansthateachthreadhandlesalldataforonesubband.
GPUs, in contrast, can efficiently exploit extremely fine 8. REFERENCES
grained parallelism. Hundreds of thousands of threads can
[1] A.R.Offringa,A.G.deBruyn,andS.Zaroubi,“Post-correlationfilter-
workinparallelonasingleGPU,withoutoverhead. Infact,
ingtechniquesforoff-axissourceandrfiremoval,” MonthlyNoticesof
in our implementation, we create one GPU thread per sam- theRoyalAstronomicalSociety,vol.422,no.1,pp.563–580,2012.
ple. We exploit data-reuse by using the shared memory that
[2] JasperJ.vandeGronde,Andre´R.Offringa,andJosB.T.M.Roerdink,
isavailableinsidetheGPU’sstreamingmultiprocessors. The “Efficientandrobustpathopeningsusingthescale-invariantrankopera-
performanceresultsareshowninfigure11. Thetoptwolines tor,” JournalofMath.ImagingandVision,vol.56,no.3,pp.455–471,
2016.
(redandgreen)showtheruntimewithoutdoinganydatain-
tegration, running SumThreshold on the full input data rate. [3] A.R.Offringa,A.G.deBruyn,S.Zaroubi,andM.Biehl, “ALOFAR
RFIdetectionpipelineanditsfirstresults,”inProc.ofScience,RFI2010,
Fortheredline,wealsoruntheSIRoperator;greeniswith-
2010,vol.II,pp.803–806.
out. Thebottomtwolinesshowperformanceifwefullyinte-
[4] JohnW.Romein,P.ChrisBroekema,JanDavidMol,andRobV.van
grateandrunSumThresholdtwotimes,onceinthefrequency
Nieuwpoort,“TheLOFARCorrelator:ImplementationandPerformance
direction,andonceinthetimedirection,asdescribedinSec- Analysis,”inProc.ACMSymposiumonPrinciplesandPracticeofPar-
tion 4. In all cases, we achieve linear Scalability. The GPU allelProgramming(PPoPP’10),2010,pp.169–178.
performssowell, thatwecanhandleallLOFARstationson [5] S.M.Ransom, S.S.Eikenberry, andJ.Middleditch, “FourierTech-
singleGPUinrealtime. niquesforVeryLongAstrophysicalTime-SeriesAnalysis,” TheAstro-
nomicalJournal,vol.124,pp.1788–1809,Sept.2002.
[6] P.ChrisBroekema, RobV.vanNieuwpoort, andHenriE.Bal, “The
squarekilometrearraysciencedataprocessorpreliminarycomputeplat-
formdesign,”JournalofInstrumentation,July2015.
6.2. IntegrationinSKAtimedomainpipeline
[7] LinusSchoemaker,“RemovingRadioFrequencyInterferenceintheLO-
FARusingGPUs,”M.S.thesis,VUUniversityAmsterdam,2015.
WearecurrentlyworkingonincorporatingourCPUandGPU
[8] P.ChrisBroekema,Albert-JanBoonstra,VictoriaCaparrsCabezas,Ton
codesintheSKACentralSignalProcessor(CSP),inthecon-
Engbersen,RobertHaas,HannoHolties,JensJelitto,RonaldP.Luijten,
text of the Time Domain Team (TDT) pulsar and transient PeterMaat,RobV.vanNieuwpoort,RonaldNijboer,JohnW.Romein,
pipeline. The initial stages of the CSP, including the beam andBertJanOffrein,“DOME:TowardstheASTRON&IBMCenterfor
former,willlikelyuseFPGAs. Afterthebeamformer,GPUs ExaScaleTechnology,”inProc.ACMworkshoponAstroHPC,2012.
are the most likely candidate for further processing. There-
fore, we will perform post-beamforming flagging using our
algorithmsonGPUs. ForanFPGAimplementation,morere-
searchisneeded.