Parallelization of Principal Component Analysis (using Eigen Value Decomposition) on scalable multi-core architecture

Parallel implementation of Principal Component Analysis(PCA) using Eigen Value Decomposition(EVD) poses many significant challenges such as Load Balancing of its modules, reducing interprocessor communication and hiding significant memory latency incurred in its modules in a manner such that the optimization can increase. It requires massive computational power while maintaining a trade-off between its numerical precision and processing time. The contribution of this paper lies in presenting an optimized parallel implementation of PCA using EVD on multi-core PowerXCell 8i without compromising on the numerical precision. It employs all the features of this architecture including SIMD vectorization, double buffering, High Memory Bandwidth and signal notification techniques. A speedup of around 40 times was achieved over single core processor for a 512×1024 matrix.