A Beowulfclass architecture proposal for realtime embedded vision

1
A Beowulf-class architecture proposal for
real-time embedded vision

• P.A. Revenga, J Sérot, J.L.Lázaro, J.P.
  Dérutin
• Dpto de Electrónica ,
• Universidad de Alcalá de Henares
• Madrid, 28806, SPAIN
• LASMEA, UMR 6602 CNRS,
• Université Blaise Pascal. Clermont-Ferrand
• Aubière, France

2
Outline

• Context
• High-performance computing
• Architecture proposal, hardware software
• SKiPPeR
• Experimental results
• Basic benchmarking results
• Realistic application, Image stabilisation
• Related Work
• Conclusion, future work

3
Context

• Real-time embedded vision
• Example assisted-driving cars, robots,
  space-rovers,
• Reactive applications
• Digital video streams I/O
• High performance (10-100fps, 0.1-5Gflops)
• Reduced volume and power consumption

4
High-performance computing

• Possible Solutions
• High-end PC.
• Specialized processors(DSP, FPGA, ASIC)
• Dedicated Multiprocessor Machines
• Beowulf-class clusters.

5
Architecture proposal

• Beowulf-class solution for real-time embedded
  vision.
• Two independent communication network
• one for interprocess comunication
• one for digital video broadcasting

6
Architecture proposalHardware

• Apple Cube motherboard
• PPC 74xx CPU
• Altivec support
• Fast-ethernet
• IEEE 1394a (400 Mbs)
• Single supply 24v
• No Fan

7
Architecture proposalHardware

• 4 - nodes
• Fast-ethernet switch
• Low power
• 4Gflops / 30 Watts
• 20x20x20 cm

8
Architecture proposalSoftware

• Layered approach
• Linux OS
• PPC 74xx with Altivec support (gcc-altivec).
• IEEE 1394 driver.
• Low level communication library.
• MPI (MPICH)
• High-level parallel programming environment.
• SKiPPER

9
Software architectureSKiPPeR

• Based upon algorithmic skeletons
• Higher-order program constructs encapsulating
  common and recurrent forms of parallelism
• Typical skeletons SCM (data-parallelism), FARM,
  
• Parallel programs built by selecting,
  instantiating and composing skeletons taken from
  a library
• Low-level implementation (mapping, scheduling,
  ..) transparently handled by the compiler

10
Software architecture SKiPPER
11
Experimental results

• Basic benchmarking
• Applying a convolution mask
• I/O on disk
• Several image and mask sizes
• Four computations times measured
• Tseq sequential ref time (1 G4 node, 450 MHz,
  128M)
• Tpar SPMD exec time (4 nodes, SCM skeleton)
• Tvec SIMD exec time (1 G4 node with Altivec)
• Tpv SPMDSIMD (4 nodes with Altivec)

12
Basic benchmarking results
13
Basic benchmarking (2)

• Relative speedups
• P/S Tseq / Tpar
• PV/V Tpv / Tvec

14
Basic benchmarking (3)

• Relative speedups
• V/S Tseq / Tvec
• PV/S Tseq / Tpv

15
Realistic application

• Image stabilisation on digital video streams
• Algorithm two stages
• Detection of Points Of Interest, with Haar
  wavelets.
• Tracking of POIs, over neighbouring windows with
  a correlation filter.
• Compute displacement
• Apply corrections

16
Image stabilisation
Search of POIs
Track
17
Image stabilisation Experimental results

• Implemented on four platforms
• Single G4 _at_ 450Mhz
• 4 x G4s with Altivec
• Single P4 _at_ 1.5Ghz
• Single G4 _at_ 1Ghz with Altivec
• Measured time to process one frame of the input
  stream
• Np nbr of POIs

18
Related Work

• Most of Beowulfs machines built today are
  dedicated to number-crunching,off-line
  computations
• Very few projects have targeted real-time
  embedded vision
• Yoshimoto et al.
• Firewire PC cluster, but using ieee1394 for data
  and video broadcast.
• NASA, REE project .
• Four node G4 processors interconnected by
  myrinet.

19
Conclusion, future work

• Originality of the work application of the
  Beowulf COTS approach to a application domain
  where dedicated solutions (hwsw) were
  traditionally used
• Good performance/Watt and Performance/dm3 ratios
• Possible applications implementation of complex
  vision and navigation algorithms on small
  autonomous vehicles, such as the CyCab