GPU Implementation of Orthogonal Matching Pursuit for Compressive Sensing

Computing outliers and related statistical aggregation functions from large-scale big data sources is a critical operation in many cloud computing scenarios, e.g. service quality assurance, fraud detection, or novelty discovery. Such problems commonly have to be solved in a distributed environment where each node only has a local slice of the entirety of the data. To process a query on the global data, each node must transmit its local slice of data or an aggregated subset thereof to a global aggregator node, which can then compute the desired statistical aggregation function. In this context, reducing the total communication cost is often critical to the overall efficiency.In this paper, we show both theoretically and empirically that these communication costs can be significantly reduced for common distributed computing problems if we take advantage of the fact that production-level big data usually exhibits a form of sparse structure. Specifically, we devise a new aggregation paradigm for outlier detection and related queries. The paradigm leverages compressive sensing for data sketching in combination with outlier detection techniques. We further propose an algorithm that works even for nonsparse data that concentrates around an unknown value. In both cases, we show that the communication cost is reduced to the logarithm of the global data size. We incorporate our approach into Hadoop and evaluate it on real web-scale production data (distributed click-data logs). Our approach reduces data shuffling IO by up to 99%, and end-to-end job duration by up to 40% on many actual production queries.

show abstract

“…Finally, note that the recovery algorithm can be accelerated with 30X ∼ 40X speed-up [2,20] using GPU which is our on-going work.…”

Section: Hadoop Implementationmentioning

confidence: 99%

Distributed Outlier Detection using Compressive Sensing

Yan

Zhang

Huang

et al. 2015

Proceedings of the 2015 ACM SIGMOD International Conference on Management of Data

View full text Add to dashboard Cite

show abstract

“…Septimus and Steinberg [2010] implemented OMP on a GPU and on FPGAs; however, this only supports images smaller than 128 elements, with a sparsity no greater than 5, which is too restrictive for radio astronomy as it only allows for images of around 11 × 11 pixels and 5 non-zero sources. Fang et al [2011] obtained a 30 to 50 times speedup for their GPU implementation of compressed sensing. Their implementation is not specialised for radio interferometric data, and thus does not take into account the algorithmic enhancements described in Chapter 4.…”

Section: Literature Reviewmentioning

confidence: 99%

Accelerated Deconvolution of Radio Interferometric Images using Orthogonal Matching Pursuit and Graphics Hardware

Belle

Armstrong

Gain

2017

J. Astron. Instrum.

View full text Add to dashboard Cite

Deconvolution of native radio interferometric images constitutes a major computational component of the radio astronomy imaging process. An efficient and robust deconvolution operation is essential for reconstruction of the true sky signal from measured correlator data.Traditionally, radio astronomers have mostly used the CLEAN algorithm, and variants thereof. However, the techniques of compressed sensing provide a mathematically rigorous framework within which deconvolution of radio interferometric images can be implemented.We present an accelerated implementation of the orthogonal matching pursuit (OMP) algorithm (a compressed sensing method) that makes use of graphics processing unit (GPU) hardware, and show significant accuracy improvements over the standard CLEAN.In particular, we show that OMP correctly identifies more sources than CLEAN, identifying up to 82% of the sources in 100 test images, while CLEAN only identifies up to 61% of the sources. In addition, the residual after source extraction is 2.7 times lower for OMP than for CLEAN. Furthermore, the GPU implementation of OMP performs around 23 times faster than a 4-core CPU.

show abstract

“…In , 8 OMP implementation on a GTX480 GPU achieved a 40x speedup over the Intel(R) Core(TM) i7 CPU. Andrecut 9 presents GPU implementation (based on the subroutines from CUBLAS) of a general Matching Pursuit algorithm, less computationally greedy than OMP.…”

Section: Introductionmentioning

confidence: 98%

“…8,9 The FPGA implementations have been done for signal length from 128-256 and sparsity values ranging from 5 to 32.…”

Section: Introductionmentioning

confidence: 99%

Parallel heterogeneous architectures for efficient OMP compressive sensing reconstruction

Kulkarni

Stanislaus

2014

SPIE Proceedings

View full text Add to dashboard Cite

Compressive Sensing (CS) is a novel scheme, in which a signal that is sparse in a known transform domain can be reconstructed using fewer samples. The signal reconstruction techniques are computationally intensive and have sluggish performance, which make them impractical for real-time processing applications . The paper presents novel architectures for Orthogonal Matching Pursuit algorithm, one of the popular CS reconstruction algorithms. We show the implementation results of proposed architectures on FPGA, ASIC and on a custom many-core platform. For FPGA and ASIC implementation, a novel thresholding method is used to reduce the processing time for the optimization problem by at least 25%. Whereas, for the custom many-core platform, efficient parallelization techniques are applied, to reconstruct signals with variant signal lengths of N and sparsity of m. The algorithm is divided into three kernels. Each kernel is parallelized to reduce execution time, whereas efficient reuse of the matrix operators allows us to reduce area. Matrix operations are efficiently paralellized by taking advantage of blocked algorithms. For demonstration purpose, all architectures reconstruct a 256-length signal with maximum sparsity of 8 using 64 measurements. Implementation on Xilinx Virtex-5 FPGA, requires 27.14 µs to reconstruct the signal using basic OMP. Whereas, with thresholding method it requires 18 µs. ASIC implementation reconstructs the signal in 13 µs. However, our custom many-core, operating at 1.18 GHz, takes 18.28 µs to complete. Our results show that compared to the previous published work of the same algorithm and matrix size, proposed architectures for FPGA and ASIC implementations perform 1.3× and 1.8× respectively faster. Also, the proposed many-core implementation performs 3000× faster than the CPU and 2000× faster than the GPU.

show abstract

GPU Implementation of Orthogonal Matching Pursuit for Compressive Sensing

Cited by 36 publications

References 7 publications

Distributed Outlier Detection using Compressive Sensing

Distributed Outlier Detection using Compressive Sensing

Accelerated Deconvolution of Radio Interferometric Images using Orthogonal Matching Pursuit and Graphics Hardware

Parallel heterogeneous architectures for efficient OMP compressive sensing reconstruction

Contact Info

Product

Resources

About