Ewan Barr: Hunting for pulsars with the SKA

Pulsars and fast radio transients are invaluable tools for probing the fundamental physics that govern our Universe. From binary pulsars facilitating incredible tests of gravity in the strong field regime, to Fast Radio Bursts potentially providing a method by which the baryon content of the Universe can be measured, it is no surprise that these phenomena are vital to many of the SKA's key science goals. However, to achieve these goals we must first expand the known populations of of these objects. In particular, if the SKA is to revolutionise our view of the Universe by opening up a new observational window in gravitational waves, we will need a large population accurately timing millisecond pulsars, the discovery of which poses large computational challenges. The extreme data rates from the SKA will require new approaches to the process of searching for pulsars and fast transients, as traditional methods simply lack the performance required. These approaches are arriving in the form of new search algorithms that use the data-parallel model offered by Graphics Processing Units and Field Programmable Gate Arrays to vastly improve the speed of our searching capabilities and enable truly real-time analysis of our data. In this talk I will introduce pulsars and fast transients in the context of the SKA and look at the challenges that we face in the coming `Big Data' era. I will also review several of the newly developed tools that may form the basis of the SKA's pulsar and fast transient search pipelines.

Transcript

1. HUNTING PULSARS WITH THE SKA Ewan Barr SKA Senior Research Fellow, Swinburne University Thursday, 27 February 14
2. OVERVIEW • What are pulsars? • Why search for them? • How do we find them? • What challenges does this present? • What is the current state-of-the-art? • Where do we go next? Thursday, 27 February 14
3. PULSARS Highly magnetised (a billion MRI machines) Neutron stars Rapidly rotating (0.12 to 716 Hz) Broadband emitters (MHz to GeV) Supra-nuclear density (humanity in a sugar cube) Extremely stable (atomic clocks in space) Thursday, 27 February 14
4. PULSARS Thursday, 27 February 14
5. PULSARS Thursday, 27 February 14
6. PULSARS Thursday, 27 February 14
7. SKA KEY SCIENCE GOALS Challenging Einstein Galaxy Evolution, Cosmology and Dark Energy Cosmic Magnetism Cosmic Dawn Cradle of Life Continuum Surveys Radio Transients Thursday, 27 February 14
8. CHALLENGING EINSTEIN: TESTS OF GRAVITY IN STRONG-FIELD REGIME Thursday, 27 February 14
9. CHALLENGING EINSTEIN: TESTS OF GRAVITY IN STRONG-FIELD REGIME R. Hulse & J. Taylor Thursday, 27 February 14
10. CHALLENGING EINSTEIN: TESTS OF GRAVITY IN STRONG-FIELD REGIME Thursday, 27 February 14
11. CHALLENGING EINSTEIN: GRAVITATIONAL WAVES Thursday, 27 February 14
12. CHALLENGING EINSTEIN: GRAVITATIONAL WAVES • Measure the ripples in space using pulsars as the arms of a Galaxy spanning interferometer. • Requires pulsars that have the most stable spins. • Benefits from an even sampling of such pulsars across the sky. • Necessitates searches for binary systems. Thursday, 27 February 14
13. SEARCH OBSERVATIONS Thursday, 27 February 14
14. SEARCH OBSERVATIONS Acquisition Thursday, 27 February 14
15. Acquisition Beam forming SEARCH OBSERVATIONS Thursday, 27 February 14
16. Acquisition Beam forming + FFT Polyphase filterbank SEARCH OBSERVATIONS Thursday, 27 February 14
17. Acquisition Beam forming Polyphase filterbank SEARCH OBSERVATIONS Search data Thursday, 27 February 14
18. SKA PHASE 1: PROPOSED SEARCH 4-bit dynamic range 50 µs sampling rate 2048 beams 0.6 Tb/s 4096 channels per sample Thursday, 27 February 14
19. PULSAR SEARCH: OVERVIEW Search in 4 dimensions: Dimension Transform N trials dependency Dispersion measure Dedispersion (Frequency)3 Acceleration TD resampling / FD template matching (Integration length)2 Pulse width Harmonic summing Minimum duty cycle Period FFT Integration length Thursday, 27 February 14
20. PULSAR SEARCH: DEDISPERSION Time Frequency t = e2 2⇡mec ✓ 1 ⌫2 a 1 ⌫2 b ◆ DM DM ⌘ Z d 0 ne dl Thursday, 27 February 14
21. PULSAR SEARCH: DEDISPERSION DDM,t = N⌫ X ⌫ A⌫,t+ t(DM,⌫) O(NtN⌫NDM) Sum all frequencies along lines of constant dispersion measure Thursday, 27 February 14
22. PULSAR SEARCH: DEDISPERSION DDM,t = N⌫ X ⌫ A⌫,t+ t(DM,⌫) O(NtN⌫NDM) Sum all frequencies along lines of constant dispersion measure Thursday, 27 February 14
23. PULSAR SEARCH: DEDISPERSION DDM,t = N⌫ X ⌫ A⌫,t+ t(DM,⌫) O(NtN⌫NDM) Sum all frequencies along lines of constant dispersion measure Thursday, 27 February 14
24. PULSAR SEARCH: DEDISPERSION DDM,t = N⌫ X ⌫ A⌫,t+ t(DM,⌫) O(NtN⌫NDM) Sum all frequencies along lines of constant dispersion measure Thursday, 27 February 14
25. PULSAR SEARCH: DEDISPERSION DDM,t = N⌫ X ⌫ A⌫,t+ t(DM,⌫) O(NtN⌫NDM) Sum all frequencies along lines of constant dispersion measure Thursday, 27 February 14
26. PULSAR SEARCH: DEDISPERSION DM = 4⇡mec⌫2 a (✏ ⌫ ⌫a)2 e2✏ ⌫(✏ ⌫ 2⌫a) = q w2 int + t2 samp + t2 DMchan Typically ~3000 trials Thursday, 27 February 14
27. PULSAR SEARCH: OVERVIEW Polyphase filterbanked beam DM trial 0 DM trial N 3000 Thursday, 27 February 14
28. PULSAR SEARCH: ACCELERATION SEARCHING • Spin frequency of pulse is Doppler shifted by motion in orbit. • Spreads signal in the Fourier domain, lowering S/N. • df/dt dependent on orbital acceleration. Thursday, 27 February 14
29. PULSAR SEARCH: ACCELERATION SEARCHING • Searching all orbital parameters is too costly. • Approximate df/dt as constant over segments of orbit. • Valid approximation for circular orbits where Tobs < Porb/10. Thursday, 27 February 14
30. PULSAR SEARCH: ACCELERATION SEARCHING • For eccentric orbits, approximation breaks down. • Either break observation and re-search, or reobserve in the hope of a better orbital phase. PSR J0737-3039 Thursday, 27 February 14
31. PULSAR SEARCH: TIME DOMAIN RESAMPLING O(NaNtNDM) Aa,t = Bt[1+a(t t obs )/2c] Stretch and compress time series to emulate frequency drift Thursday, 27 February 14
32. PULSAR SEARCH: TIME DOMAIN RESAMPLING a = 48 c t2 obs s✓ 1 ✏4 1 ◆ Ntrials depends on tobs 2 (10 mins gives 700 trials) Thursday, 27 February 14
33. PULSAR SEARCH: OVERVIEW Polyphase filterbanked beam DM trial 0 DM trial N ACC trial 0 ACC trial N ACC trial 0 ACC trial N 3000 2100000 Thursday, 27 February 14
34. PULSAR SEARCH: FAST FOURIER TRANSFORM O ( NaNDMNt log2 Nt) Best performance with prime factorable N Real to complex FFT, exploits Hermitian symmetry to reduce complexity Thursday, 27 February 14
35. PULSAR SEARCH: SPECTRAL INTERPOLATION DFT response is imperfect at bin edges Interpolate to improve response to arbitrary frequencies Thursday, 27 February 14
36. PULSAR SEARCH: SPECTRAL INTERPOLATION O(NaNDMNt) Ai = max ✓ Bi, 1 p 2 ( Bi + Bi+1) ◆ Thursday, 27 February 14
37. PULSAR SEARCH: HARMONIC SUMMING • Pulse power spread in Fourier domain. • Incoherently add harmonics to increase signal. • For Nh harmonic numbers. Thursday, 27 February 14
38. PULSAR SEARCH: HARMONIC SUMMING / PEAK FINDING Ai,Nh = 1 p Nh Bi + Nh X h B(ih/Nh) ! O(2Nh NaNDMNt) • After each harmonic sum we threshold the spectrum and mark candidates above the threshold. • Sort candidates above threshold by signal-to-noise or power. • Store candidates for application of clustering algorithms. O(NhNaNDMNt) Thursday, 27 February 14
39. PULSAR SEARCH: OVERVIEW Polyphase filterbanked beam DM trial 0 DM trial N ACC trial 0 ACC trial N ACC trial 0 ACC trial N Harm trial 0 Harm trial N Harm trial 0 Harm trial N Harm trial 0 Harm trial N Harm trial 0 Harm trial N 3,000 2,100,000 10,500,000 Thursday, 27 February 14
40. PULSAR SEARCH: OVERVIEW Clustering Algorithm Harm trial 0 Harm trial N Harm trial 0 Harm trial N Harm trial 0 Harm trial N Harm trial 0 Harm trial N 10,500,000 Candidate frequencies Unique candidates Thursday, 27 February 14
41. PULSAR SEARCH: OVERVIEW Clustering Algorithm Harm trial 0 Harm trial N Harm trial 0 Harm trial N Harm trial 0 Harm trial N Harm trial 0 Harm trial N 10,500,000 Candidate frequencies Polyphase filterbanked beam Folding Algorithm Candidate Fold 0 Candidate Fold 1 Candidate Fold 2 Candidate Fold 3 Candidate Fold 4 Candidate Fold N ~1000 Thursday, 27 February 14
42. PULSAR SEARCH: FOLDING O(NcandN⌫Nt) Thursday, 27 February 14
43. PULSAR SEARCH: CLASSIFIERS Candidate classifiers + = Thursday, 27 February 14
44. Thursday, 27 February 14
45. SKA PHASE 1: EXAMPLE SEARCH Dedispersion (3000 trials) 0.5 Eops Resampling (700 trials) 0.1 Eops FFT 2 Eops Spectrum forming 0.1 Eops Harmonic summing 1.5 Eops Peak finding 0.1 Eops Folding 50 Pops Total ~4.3 Eops Thursday, 27 February 14
46. SKA PHASE 1: EXAMPLE SEARCH ~7 Petaflops Thursday, 27 February 14
47. WHAT TO USE? • Pulsar astronomers are a fickle bunch... • Hardware should preferably be flexible enough for algorithmic change to be easily implemented in a short timeframe. • Hardware must be fast... • GPUs, FPGAs and ASIC are the main competitors. • GPUs: Flexible but high power consumption. • FPGAs: Low power consumption but relatively inflexible. • ASIC: Low unit cost but high development cost and no flexibility. Thursday, 27 February 14
48. WHAT DO WE HAVE NOW? • Several tools for dedispersion: • ARTEMIS GPU dedispersion (Wes Armour) • Dedisp GPU dedispersion (Ben Barsdell) • TARDIS FPGA dedispersion (Nathan Clarke) • One (and a half) tools for everything else: • Peasoup end-to-end GPU pulsar search (Me) • PRESTO GPU accelsearch (Jintao Luo & Scott Ransom) • We still need an optimal GPU folding algorithm Thursday, 27 February 14
49. WHAT’S NEXT? Survey for Pulsars and Extragalactic Radio Bursts (SUPERB) • Mid-to-high Galactic latitude search using Parkes shadowed by Molonglo. • 0.77 Gb/s. • We will do real-time transient and real-time (mildy) accelerated pulsar searching using Peasoup. • Will use 3 7-GPU nodes of the gSTAR cluster (Fermi C2070s) Thursday, 27 February 14
50. WHAT’S NEXT? Survey for Pulsars and Extragalactic Radio Bursts (SUPERB) Look out for exciting new results! Thursday, 27 February 14