Portable and Predictable Performance on Heterogeneous Embedded Manycores (ARTEMIS ) ARTEMIS 2 nd Project Review October 2014 Aerospace Demonstrator Ricardo Moreno (TAS-E) ARTEMIS PaPP Review 20131
Contents Rationale Hardware Platform Use case: algorithm CCSDS 122 Parallelization with OpenMP Results and demonstration Contribution to PaPP objectives Next year goals ARTEMIS PaPP Review 20142
Rationale Continuous growth of quality and size of images taken from satellite but, limited downlink bandwidth due to technological limitations Solution: Compression of on-board data prior to transmission to ground At the same time: Deal with strong dependability and safety constraints Multi-client and tedious certification process System deterministic behaviour Goal: increase performance and maintain quality ARTEMIS PaPP Review 20133
Hardware platform: multicore Leon ARTEMIS PaPP Review 20144
Algorithm CCSDS 122: overview Payload lossy and lossless data compressor for bidimensional images instruments and potentially for multispectral and hyperspectral imagers and sounders Two parts: DWT + BPE ARTEMIS PaPP Review 20145
Algorithm CCSDS 122: OpenMP benefits Shared memory: Pthread and OpenMP preferred against MPI or MCAPI OpenMP preferred: Lower modification of original source Easier to synchronize tasks Hybrid Task/Data parallelism solution: Split two tasks: DWT + BPE Data parallelism exploit within each task Pipelining not implemented due to unbalanced execution time on each task ARTEMIS PaPP Review 20146
Parallelization with OpenMP: First approach ARTEMIS PaPP Review FIR ¼ DWT
Each 1d DWT iteration is parallelized by using OpenMP paradigm Based on pragmas (for loops) DWT example: ARTEMIS PaPP Review Parallelization with OpenMP: First approach
Parallelization with OpenMP: Results I First test in x86 platform Intel 3770K Quad-core No BPE parallelization Real images from satellite: ARTEMIS PaPP Review 20149
Parallelization with OpenMP: Results II Amhdal’s law ARTEMIS PaPP Review
Demonstrator: PaPP development system ARTEMIS PaPP Review
Image generation ARTEMIS PaPP Review
TAS-E board : Quad-core Leon3 SoC synthesized in FPGA Real target attached to development PC ARTEMIS PaPP Review
Image download and application execution ARTEMIS PaPP Review
Contribution to PaPP objectives Primary objective 2: Portability of performance across at least two hardware platforms for the application use cases. Portability from x86_64 to Leon platform Primary objective 3: Portability of the software stack across application domains. Portability to aerospace domain Primary objective 4: Software developer productivity is increased OpenMP requires lower programmer skills compared to other parallel programming paradigms ARTEMIS PaPP Review
Next year steps Parallelization of bpe (second part of algorithm) Use of OpenMP tasks (currently for-loops parallelization) Better integration with WP3 tools for performance predictability Evaluation of results ARTEMIS PaPP Review
ARTEMIS PaPP Review