Data Intensive Computing
Hybrid Computing Solutions Applied to Feature Extraction, Characterization, Classification, and Clustering
Principal Investigators: Harold Trease, Rob Farber, Adam Wynne, Lynn Trease
Challenge
- Testing new architectures and programming models for hybrid multi-core computing.
- Weather low power high perf commodity processors can be used for real applications and to see if we can use current sw env or need new one for hw - right now the proces are avial, but special purpose, other apps like image processing for video analysis, but we have to wrtie those apps ourselves,
- People don't do that level of real time processsing, on procesor computing; many surveill camers, data capture stored but not reviewed until forenscis
- Realtime predictive
- Cheap, available
- Scoping out applicability of current sw tools to new hw archs
- What would design of next gen be?
Approach
Multi-core hardware:
- Linux cluster, Cell cluster, Nvidia GPGPU cluster, Cray XMT
Programming models:
- MPI, Hadoop, Star-P, MeDICI
Impact
Showing the value, flexibility and (extreme) performance of hybrid multi-core processing
Collaborations
- TeraGrid high performance computing (Texas Advanced Computing Center, Argonne National Laboratory)
Accomplishments
- MPI Control network for communicating between multi-core machines
- Excellent scaling of Nvidia GPGPU performance (~200 x X86 results → 500GFlops/sec)
Feature Extraction, Characterization, Classification, and Clustering