A real-time software backend for the GMRT : towards hybrid backends CASPER meeting Capetown 30th September 2009 Collaborators : Jayanta Roy (NCRA) Yashwant Gupta (NCRA) Yashwant Gupta (NCRA) Ue-Li Pen (CITA) Jeff Peterson (CMU) Jitendra Kodilkar (NCRA) Sanjay Kudale (NCRA)

Existing GMRT backend  30 antennas, 32 MHz, dual pol, correlator + beam former  Based on ASICs and DSPs ; designed in early 1990s ; running for more than 10 yrs now  Limitations :  Lack of flexibility in time and frequency resolution  Limited number of bits  Obsolescence; difficult to maintain  New, desirable features :  Capability for recording raw voltage signals  Facility to add new algorithms e.g. RFI mitigation  Ability to form multiple beams within the primary beam

Software backends Software based back-ends : Software based back-ends : Few made to order hardware components ; mostly off-the-shelf items Few made to order hardware components ; mostly off-the-shelf items Easier to program ; more flexible Easier to program ; more flexible GMRT Software Back-end (GSB) : GMRT Software Back-end (GSB) : 32 antennas 32 antennas 32 MHz bandwidth, dual pol 32 MHz bandwidth, dual pol Net input data rate : 2 Gsamples/sec Net input data rate : 2 Gsamples/sec FX correlator + beam former FX correlator + beam former Uses off-the-shelf ADC cards, CPUs and switches to implement a fully real-time back-end Uses off-the-shelf ADC cards, CPUs and switches to implement a fully real-time back-end Raw voltage recording to disks, for all antennas; off-line read back & analysis Raw voltage recording to disks, for all antennas; off-line read back & analysis Currently status : completed and ready to release as observatory facility (from 15 th October) Currently status : completed and ready to release as observatory facility (from 15 th October)

The GMRT software backend : block diagram

The ADC Card  PCI-X compliant ADC card (Acquisition Logic)  4 channels, 8 bits each  External trigger (from GPS pulse)  External reference clock (from Rb)  Several operating modes :  33 / 66 MHz sampling, 4 channels (8 / 4 bits)  100 MHz, 2 channels (4 bits)  200 MHz, 1 channel (2 bits)  Data transfer from ADC card to memory using 145 MB/s

Software flow : real-time mode PABeam IABeam ADC 16 MHz or 32 MHz (withAGC)IntDelayCorrectFilter+DesampFFT+FSTC&Fringe MAC Beamformer visibilities 64 analog Inputs (32 ants, 2 pols)

Software flow : real-time mode

Software flow : raw voltage mode

Performance Optimisation  Network transfer optimisation : jumbo packets  Computation optimisation :  Intel IPP routines (for FFT)  Vectorised operations  Cache optimisation  Multi-threading load balancing  Performance specs :  Better than 85% compute efficiency  $190 / baseline ; 250 Mflops / W

Sample Results : Imaging  J calibrator field at 1280 MHz  8.5 hrs synthesis image  Central source : 4.83 Jy  Noise level at HPBW : 34 microJy  Dynamic range achieve : ~ 1.5 x10 5

Sample Results : Beamforming  Phasing the array using a point source calibrator  Single pulses from PSR B

Sample Results : Coherent Dedispersion

Real-time Data Monitoring

New Capabilities : RFI mitigation  MAD filtering on raw time resolution data to eliminate bursty, time domain RFI : works very nicely

Future Plans : Multiple Beamforming Offline processing mode of software back-end : beam-forming to produce multiple phased array beams covering the primary beam – for pulsar and transient science Offline processing mode of software back-end : beam-forming to produce multiple phased array beams covering the primary beam – for pulsar and transient science

Multiple Beamforming  Can get full phased array sensitivity over the entire field of view  Useful for searching for pulsars in extended targets (e.g. SNRs, GCs)

New Plans : Transient Search Pipeline  Collaborative effort of NCRA and Swinburne University of Technology  Running in piggy-back mode simultaneously with other observations  Object of interest : Fast transients -- nanosecond to 100's of millisecond  Event detection : based on the sensitivity of 8 antennae incoherent array beamover 32 MHz  Search in Dispersion measure space : Discriminate out fast radio transient from RFI.  Coincidence or anti-coincidence filter : Multiple sub-array multiple beam coincidence filter reduces the false triggers due to direction- location dependent RFI.  Efficient RFI removal technique

New Plans : Transient Search Pipeline

Multiple incoherent beam formation for transient pipeline

3 incoherent array beams : PSR B Detection of RFI by coincidence filter RFI ? Beam-1 Beam-2 Beam-3

Future Possibilities : Towards Hybrid Designs  Upgraded GMRT : 32 stations, 400 MHz BW, dual pols  ~ 200 nodes (8 cores each)  GPU accelaration feasible  Mix and Match of FPGAs and CPUs/GPUs  Software backend for special applications :  Spectral zoom modes  Raw voltage recording modes  Debugging mode for new arrays

Thank You