high performance streaming data processing
play

High Performance Streaming Data Processing F. Huizinga Radio - PowerPoint PPT Presentation

High Performance Streaming Data Processing F. Huizinga Radio Astronomy Study celestial objects in radio frequencies LOFAR ( LO w F requency AR ray) 7000 Antennas Below 250 MHz Multiple stations Biggest in the world


  1. High Performance Streaming Data Processing F. Huizinga

  2. Radio Astronomy ● Study celestial objects in radio frequencies ● LOFAR ( LO w F requency AR ray) ○ 7000 Antennas ○ Below 250 MHz ○ Multiple stations ○ Biggest in the world ○ New design

  3. Example image LOFAR

  4. LOFAR Superterp ● 6 Stations ● 288 usable ants ● 300m diameter

  5. New Antenna Design ● Direction by phase delays between antennas ● Highly computational ● Cheap antennas

  6. Enter The AARTFAAC Project ● A msterdam A stron R adio T ransients F acility A nd A nalysis C entre ● 120 degrees FOV ● Realtime imaging ● Looking for transients ○ Non static objects ○ Pulsars ○ Galaxies ○ Jets ● Superterp only ● Trigger LOFAR

  7. Computational Challenges ● Correlator: 12 Uniboards of FPGAs @ Astron ● Visibilities: (288*289)/2*4 = 166464 (dual pol.) ● Bandwidth: up to 13.8 MHz (576 images) ● Dump time: Every 1/10th second ● Data rate: 40 Gb/s ● Commensal to LOFAR ● Responsetime: <= 1 second

  8. Pipeline Framework Candidates ● Hadoop ○ - No streaming data (batch processing) ● Storm ○ + Twitter ○ + Define your own topology easily ○ - Meta data overhead ● Pelican ○ + Close collaboration ○ + Astronomy examples/projects ○ - Young ● 0-MQ ○ - Low level ○ + Large userbase, docs

  9. Pelican Framework (1/2) ● Streaming data processing ● Developed by OeRC (Oxford) ● C/C++ ● Depends ○ Boost (minimal) ○ Qt framework ○ CppUnit ○ Doxygen ○ CMake ● Designed with astronomy in mind ● Modular

  10. Pelican Framework (2/2) ● Emulator Emulator ○ Read from disk, self ○ Send udp packets Pelican Server ● Server (bottleneck) ○ Creates chunks from udp ○ Sends tcp packets Pipelines Module 1 ○ First come first served Module 1 Module 1 ● Pipeline Module 2 Module 2 Module 2 ○ Creates datablobs from tcp Module N Module N ○ Manipulate datablobs with Module n modules ○ Send/Store elsewhere

  11. Antenna Pipeline Design Antenna Antenna ● Matlab Correlator ● Pelican Server Calibrator ● CppUnit Monitoring and Control ● Eigen3 Pipeline 1 Pipeline 2 Pipeline n ● Jenkins Flagger Flagger Flagger Calibrator Calibrator Calibrator Imager Imager Imager Image cube Transient Pipeline

  12. Example Image AARTFAAC

  13. More info ● Sourcecode ○ Sorry, not yet ○ Nearby future!! ○ Some legal issues ● www.aartfaac.org ● www.lofar.org ● github.com/pelican/pelican ● eigen.tuxfamily.org

  14. Questions ?

Recommend


More recommend