1. Ben Barsdell Matthew Bailes Christopher Fluke David Barnes

2. Background  High Time Resolution Universe (HTRU) survey  Uses the Parkes 64m radio telescope  Goal is to discover new pulsars and radio transients Survey specs 400 MHz BW @ 1381.8 MHz 1024 freq. channels 64μs time resolution 2-bit sampling

3. Event Multi-beam receiver Incoherent dedispersion Dedispersed time series DM Time Signal search List of candidates Masked data Freq. Time RFI removal Parkes Filterbank data Freq. Time Swinburne Follow-up observation Ben Barsdell - ADASS 2011

4. The Plan  Current pipeline takes > 30 mins per 10 min observation  Necessitates off-line processing  Means transfers, tapes and long waits  Would like to speed things up to real-time  Instant feedback and follow-up observations  Triggered baseband data dumps  How to do it?  Time to bring out the heavy artillery… Ben Barsdell - ADASS 2011

5. The GPU Ben Barsdell - ADASS 2011

6. Event Telescope receiver beams Filterbank data Freq. Time Incoherent dedispersion Dedispersed time series DM Time Signal search List of candidates Masked data Freq. Time RFI removal Parkes Swinburne Follow-up observation Ben Barsdell - ADASS 2011

7. The detection pipeline Incoherent dedispersion RFI mitigation Baseline removal Sigma clip Report candidates RFI mitigation Fourier transform Fourier search Harmonic summing Matched filter Single pulse search Ben Barsdell - ADASS 2011

8. RFI mitigation  Interference is a big problem  No easy solution  Military radar too important  Prime-time TV too popular  Some things can be done  Sigma clipping  Spectral kurtosis  Coincidence rejection www.clker.com Ben Barsdell - ADASS 2011

9. Coincidence rejection  Use multi-beam receiver as reference antennas  Assume RFI is not localised  Apply simple coincidence criteria:  E.g., 3σ in 4+ beams => RFI  Or use Eigen-decomposition approach  Run on GPU as a straightforward transform  RFI_mask[i] = is_RFI(multibeam_data[i])  Note: Eigen-decomp method makes is_RFI() trickier Ben Barsdell - ADASS 2011

10. Dedispersion Radio source Broadband signal Frequency (MHz) Phase Dispersed signal Phase Frequency (MHz) Ben Barsdell - ADASS 2011  Unknown distance => search through DM space  Pick DM, dedisperse, search, repeat  ~ 1200 DM trials e-e- ISM ?

11. Dedispersion  Computationally intensive problem  Biggest time-consumer  Runs really well on a GPU  Lots of parallelism  High arithmetic intensity  Good memory access patterns  No branching Ben Barsdell - ADASS 2011

12. Other algorithms  Baseline removal  Subtract running mean  Port to GPU using parallel prefix sum  Matched filtering  Convolve with 1D boxcar  Can also use parallel prefix sum  Sigma cut + peak find  Threshold and segment  Port to GPU using segmented reduction Ben Barsdell - ADASS 2011

13. GPU dedispersion Ben Barsdell - ADASS 2011

14. Preliminary results  Dedispersion: 20 mins  < 2.5 mins  Using ‘direct’ method on 1 Tesla C2050 GPU  Time-binning gives further 2x speed-up  Details in Barsdell et al. 2012  Other algorithms mostly complete  1 beam / GPU should be easy  2 beams / GPU within reach? Ben Barsdell - ADASS 2011

15. Software/hardware configuration Ben Barsdell - ADASS 2011 1 2 3 4...... Recv beamDedisperseSend DMsRecv beams Send DMs.................. Continue…...... ProcessOperation

16. Deployment  Destined for Parkes  Real-time results  RFI ‘weather report’ Ben Barsdell - ADASS 2011

17. Looking ahead  Real-time radio transient detection promises to  Simplify the data processing procedure  Enable immediate follow-up observations  Allow capture of high-resolution baseband data for significant events  Catch things like the ‘Lorimer burst’ as they happen! Ben Barsdell - ADASS 2011