Porting CPU-based Multiprocessing Algorithms to GPU for Distributed Acoustic Sensing

Transcription

1 GTC2014 S4470 Porting CPU-based Multiprocessing Algorithms to GPU for Distributed Acoustic Sensing Steve Jankly Halliburton Energy Services, Inc.

2 Introduction Halliburton Halliburton is one of the world s largest providers of products and services to the energy industry. Pinnacle Pinnacle provides industry-unique integration of various fiber optic technologies for fracture diagnostic and reservoir monitoring Fiber optic based sensing technologies can be highly data-intensive 2

3 Fiber Optic Sensing Fiber optic sensing has been shown as a beneficial replacement of conventional sensors A fiber optic cable may be used as an actual sensor, rather than being used for transmitting data. 3

4 Fiber Optic Sensing Various technologies for optical sensing exist: DAS DTS FBG And more 4

5 Distributed Acoustic Sensing (DAS) What is DAS? DAS turns your run-of-the-mill fiber optic cable into a fully distributed sensing system When applied to an oil and gas well, operators can tune in to any section of the fiber and see what is going on Applications are varied: Hydraulic fracture stimulation and production monitoring Verification of downhole equipment operation Pipeline integrity Collection of seismic imaging data 5

6 Distributed Acoustic Sensing (DAS) How does DAS work? Sends light pulses into a fiber optic cable The light is backscattered in the fiber, which is then processed at the receiver end This generates a large amount of data which must be processed in real-time We have developed two DAS Technologies: DAS-I: Intensity based DAS interrogation Simple data processing DAS-P: Phase based DAS interrogation More complicated algorithm: Involves (Phase) Demodulation 6

7 DAS-Intensity Basic design: Provides real-time intensities along the fiber Depending on the fiber length and other parameters, there can be up to thousands of channels of data FPGA/DSP data processing / hardware control Gigabit Ethernet data interface 7

8 DAS-Intensity Data output Custom protocol over TCP Command interface: supports configuration, calibration, and data acquisitions Instrument is the server, and applications calling the instrument API are clients An API is provided, which is a black box around the custom protocol 8

9 DAS-Intensity Benefits Applications where only intensity data is needed Low power due to embedded processing (FPGA/DSP) Remote usage (Ethernet) Limitations Low processing capability of the FPGA/DSP combo DSP has only 1 core Throughput is limited by Gigabit Ethernet Cannot be upgraded to DAS-P What is the solution? 9

10 DAS-Phase We quickly found that GPU computing is the answer Why did we not utilize an embedded solution this time? DAS-P is more memory and processing intensive: Large memory requirements Optical data processing algorithm API was developed to: configure system elements acquire and process data 10

11 DAS-Phase System Design Hardware design expanded upon DAS-I: Optics modified to provide additional information about the optical signal Support for increased data rates Instrument is paired with a customized PC that digitizes and processes the raw data stream 11

12 CPU/GPU Data Processing Algorithm implementations CPU (Intel Core i7) Takes advantage of multiple CPU cores Necessary if GPU is not available, such as in a scaled down version of the product GPU (NVIDIA Tesla) The data processing algorithm was ported to NVIDIA GPUs, utilizing CUDA C API supports selective processing on CPU or GPU 12

13 CPU/GPU Data Processing CPU/GPU Data Flow Raw data in at full resolution can be up to ~800MB/sec The digitizer runs at 100MSPS and future versions may even be greater The optical data processing algorithm provides more information about the fiber optic signal Phase and amplitude information 13

14 CPU/GPU Data Processing CPU/GPU Data Flow 14

15 CPU/GPU Data Processing Porting algorithm from CPU to GPU Conceptually straightforward to port to GPU since algorithm is parallelizable Same algorithm is applied to many data points Improved various stages of CPU-based algorithm as a side-effect after the code refactoring process for GPU API was modified to manage the memory and spawn threads in GPU, if GPU processing is selected 15

16 CPU/GPU Data Processing Speedups over non-optimized CPU CPU basis of comparison: Intel Core i (3.60 GHz, 4-core) Optimized CPU algorithm was tested Two GPUs were tested: NVIDIA Tesla K20c NVIDIA Quadro K5000M 16

17 CPU/GPU Data Processing Overall speedups against non-optimized CPU: GPU and Optimized CPU Implementation Speedups K20C K5000M CPU_opt Speedup 17

18 Conclusion GPU acceleration allowed DAS-P to support higher sensing resolutions and ranges in real-time More powerful and/or additional GPUs will enable next generation products Future Plans Further optimize memory usage and algorithm Possible next gen systems: Consolidate DAS-P into one chassis Develop new custom embedded board Increase data rate Possible alternative data output interfaces 10 gigabit Ethernet 18

19 Questions? 19