escience: Pulsar searching on GPUs

Size: px

Start display at page:

Download "escience: Pulsar searching on GPUs"

Samantha Henderson
5 years ago
Views:

1 escience: Pulsar searching on GPUs Alessio Sclocco Ana Lucia Varbanescu Karel van der Veldt John Romein Joeri van Leeuwen Jason Hessels Rob van Nieuwpoort And many others!

Netherlands escience center Science itself is changing We need to change with it escience = Enhanced science Apply ICT in the broadest sense Data-driven research

2 Netherlands escience center Science itself is changing We need to change with it escience = Enhanced science Apply ICT in the broadest sense Data-driven research across all scientific disciplines Develop generic escience tools Collaboration between disciplines Coordinate escience approaches Contribute to scientific breakthroughs

3 Powered by SURF and NWO Also a funding organization Currently Big Data call escience integrators Top researchers in a variety of disciplines Form bridge between domain and escience escience engineers The innovators that do the real work

and Cogntion/Neuroscience) 3. escience Methodology & Big Data (Astronomy) 4.

4 Priority areas 1. Chemistry & Materials 2. Life Sciences (Green Genetics, Life Sciences & Translational Research, Foods and Cogntion/Neuroscience) 3. escience Methodology & Big Data (Astronomy) 4. Humanities & Social Sciences 5. Sustainability & Environment (Climate, Water management, Energy and Ecology)

5 Big Data Many novel instruments and applications produce lots of data LHC, telescopes, climate simulations, genetics, medical scanners, facebook, twitter, These applications cannot work without computer science anymore A lot of unexploited knowledge that we can only find if the data across disciplines is accessible and usable Challenges: data handling, processing Need large-scale parallelism

6 Searching for pulsars Rotating neutron starts Discovered in 1967 About 2000 are known Big mass, precise period Lab in the sky Probe space and gravitation Investigate general relativity Work done by Alessio Sclocco

7 Pulsar pipeline Real-Time Central Processing Poly-Phase Filter RFI mitigation Beam Former Dedispersion Folding Signal-to-Noise Pulsar Pipeline

8 Portability and Flexibility Many different instruments Software telescopes Life time of an instrument is much longer than life time of compute hardware Big Data, Big Compute and heterogeneity DAS-4 is ideal research platform

9 OpenCL: The Khronos group

10 OpenCL: Open Compute Language Architecture independent Explicit support for many-cores Low-level host API Uses C library, no language extensions Separate high-level kernel language Explicit support for vectorization Backend for high-level models: Pieter Hijma

11 Performance portability Many different observation types and parameters Only known at run time E.g. # frequency channels, # stations, longest baseline, filter quality, observation type Use runtime compilation and auto-tuning Map specific problem instance efficiently to hardware Auto tune platform-specific parameters Portability across different platforms / families

12 Pulsar pipeline Real-Time Central Processing Poly-Phase Filter RFI mitigation Beam Former Dedispersion Folding Signal-to-Noise Pulsar Pipeline

13 Beam forming Apply correct delays Combine stations Weighted addition of all stations Alessio Sclocco [IPDPS 2012]

14 Auto-tuning GTX580 OpenCL

15 Auto-tuning GTX580 CUDA

16 Beam forming performance GFLOPS IBM Blue Gene/P Intel Xeon E5620 (OpenMP/SSE) Intel Xeon E5620 (OpenCL) AMD Opteron AMD Opteron (OpenCL) (OpenMP/SSE) AMD HD6970 (OpenCL) NVIDIA GTX580 (CUDA) NVIDIA GTX580 (OpenCL)

17 Beam forming power efficiency GFLOPS / Watt IBM Blue Gene/P Intel Xeon E5620 (OpenMP/SSE) Intel Xeon E5620 (OpenCL) AMD Opteron 6172 (OpenMP/SSE) AMD Opteron 6172 (OpenCL) AMD HD6970 (OpenCL) NVIDIA GTX580 (CUDA) NVIDIA GTX580 (OpenCL)

18 Pulsar pipeline Real-Time Central Processing Poly-Phase Filter RFI mitigation Beam Former Dedispersion Folding Signal-to-Noise Pulsar Pipeline

19 Real-time RFI mitigation Some pipelines run in real time pulsar searching RFI is much stronger than pulsars Beam forming takes union of RFI of all receivers

20 Real-time RFI mitigation challenges Limited amount of samples Only 1 second, no data from the future, only statistics from the past Fewer frequencies Only one frequency band per node Each second, a node gets a different frequency Real time We can afford only few operations per byte

21 RFI mitigation

22 Pulsar pipeline Real-Time Central Processing Poly-Phase Filter RFI mitigation Beam Former Dedispersion Folding Signal-to-Noise Pulsar Pipeline

23 Pulsar pipeline DM Three unknowns: Location: create many beams on the sky Dispersion: measure of distance of pulsar Period Brute force search across all parameters Everything is trivially parallel (or is it?) Our approach: keep everything in the GPU period

24 Dedispersion Frequencies have different delays through ISM Dispersion measure (DM) Brute-force search Output has no channels anymore

25 Data reuse

26 Performance on GPUs

27 Period search: Folding Stream of samples Period 8: Period 4:

28 Period search: Folding Build a tree of periods to maximize reuse Each thread block computes a path from leaf to root

29 Signal-to-noise SNR = (max mean) / RMS Computed on folded data Run time negligible compared to dedispersion and folding Output a list of candidates Reduced terabytes to only 15 numbers

30 Pulse profile Signal power time

31 Conclusions Cuda / OpenCL performance gap gone Portable: OpenCL and auto-tuning Entire pipeline on GPUs No PCI-e bottleneck moving data back and forth Trivially parallel? Yes, but not trivial! It s all about the memory escience approach works! Requires commitment from both sides

Signal Processing on GPUs for Radio Telescopes

Signal Processing on GPUs for Radio Telescopes John W. Romein Netherlands Institute for Radio Astronomy (ASTRON) Dwingeloo, the Netherlands 1 Overview radio telescopes motivation processing pipelines signal-processing