Optimising In- Field Processing using GPUs Tarik Saidani Senior - - PowerPoint PPT Presentation

Optimising In-Field Processing using GPU’s

Tarik Saidani

Senior Software Engineer, PGS

Peng Wang

DevTech, Nvidia

From a Seismic Acquisition Survey

To a High Resolution Image of The Sea Subsurface

Problem: Source and Receiver Ghost

sea surface source far field direct ray-path direct ray-path ghost ray-path ghost ray-path cable (receiver) receiver ghost source ghost

A Ghost Free Marine Acquisition System

5

• Method

– Combine pressure and velocity sensors in a solid streamer – Use the complementary ghost patterns of the two sensors to remove the receiver ghost – Tow the dual sensor streamer deep for low frequency content

• Result

– The bandwidth of the data is increased for both low and high frequencies when compared with conventional streamer data – There is better low frequency penetration

• f the source signal

– The acquisition method is less sensitive to weather conditions

Solution: Dual Sensor Streamer Acquisition

6

The Big Data Challenge In the Seismic Business

1995 6 streamers 2005 16 streamers 2015 24 streamers

Seismic Acquisition Data Volumes

• A typical streamer is 8000 meters long and contains 1280 receivers
• Data is recorded in time chunks or as continuous series, 2ms sample interval,

generating 500 samples per second per receiver

• A streamer (single sensor) generates 640,000 samples per second
• One streamer spread (10 streamers) generates 6,400,000 samples per second
• Big spread, 20 streamers, dual sensor will generates 25,600,000 samples per

second

• Typical acquisition can generate multi-TBs of data per day

3D Wave-field Separation Workflow

9

upsampling Receiver Deghosting 84 % 16% 100 %

Getting the Best Possible Image from the Early Stages

10

Streamer-wise Wavefield Separation 3D Wavefield Separation

Upsampling

• Iterative process
• Frequency domain (wavenumber)
• Not enough parallelism in inner loop (few thousands of threads)
• Window parallelism not exposed in the CPU code
• Loop restructuring to expose window parallelism
• After the code change enough parallelism for the GPU (millions of threads)

11

Receiver Deghosting

• Large volume of data (hydrophone and geosensor data)
• Frequency domain computations
• Parallelism over traces and frequency samples
• Fairly straightforward parallel code
• Parallelism available at many loop levels on a large number of iterations

12

Infield Constraints

13

• Although she looks big in the

picture the ship has very limited space to host a compute cluster

• Power and cooling are also

limited on-board a vessel

• A CPU based solution was

considered but was quickly discarded because of the constraints described above

But Also Facing Up to New Realities …

Phase 1: Getting the Most of the CPU Cycles

• CPU code profiling and analysis
• Hotspot analysis showed that not much could be improved
• The vectorizer was not doing a great job, had to write vector intrinsics …
• Reached an upper bound in terms of CPU performance … not enough!

15

Phase 2: What Can We Do Next?

• Parallelism already present at different levels: thread, process, vectorization …
• We can not rely on increasing the CPU core count because of the above

constraints

• GPU accelerators were the most obvious way forward
• GPU prototype code:

– Ported the streamerwise degohsting code to the GPU – 25x speedup compared to the single core CPU (Haswell) – 7x speedup on the entire flow: interesting …

16

Phase 3: The Bigger Picture

• The streamerwise deghosting code having been ported to the GPU, upsampling

was the new hotspot in the flow

• Two parallel development branches:

– Porting the upsampling code to the GPU – Porting the 3D deghosting code to the GPU

• At the end of this phase the processing flow was 15x faster
• In the meantime an additional processing step was added to the deghosting code:

extrapolation

• It increased the runtime and changed the application profile (50% of the runtime).

17

Phase 4: Putting it All Together

• Ported the extrapolation to the GPU
• Very similar compute kernel to the upsampling
• The first benchmarks showed a throughput that was 40x faster than 1 CPU core
• After running more production like tests we achieved an impressive 100x!

18

Hardware Footprint

19

CPU based GPU based

20:1 Nvidia Tesla K80

Summary

• Wavefield separation is a fundamental step in marine data acquisition and processing
• It is a very demanding process in terms of compute power
• Infield constraints discard large scale systems
• In order to deliver an acceptable throughput within an acceptable footprint the only

viable solution is GPU based

• The final result showed an impressive throughput along with a very small footprint
• It Improves the geophysical quality of PGS field acquisition deliverable
• Real-time 3D processing of data during acquisition
• GPU deployment started on vessels in Q1 2016

20

Titan Class Tethys, Now With GPU-Based “3D Wavefield Separation Appliance”

Acknowledgment

Peng Wang Ty Mckercher Ken Hester

22