Atlas Tracking Optimization on GPU Luis Domingues Professor: - - PowerPoint PPT Presentation

Master Thesis

Atlas Tracking Optimization on GPU

Luis Domingues

Professor: Frédéric Bapst Supervisors: Paolo Calafiura Wim Lavrijsen Expert: Mathieu Monney

02/25/2015

Luis Domingues - January 2015 2

Target

Luis Domingues - January 2015 3

Code we started from

• Demonstrator of ATLAS trigger on GPUs
• Basic host side

– Take data – Send and compute data on GPU – Sleep waiting the response

Luis Domingues - January 2015 4

Code we started from

Luis Domingues - January 2015 5

Overlapping pixels and SCT

• The pixel and SCT processing are done in sequence
• Same event, but sequential processing...

Time Pixel SCT Kernels Time stamp Time stamp Kernels Time stamp Time stamp

Luis Domingues - January 2015 6

Overlapping pixels and SCT

Luis Domingues - January 2015 7

CUDA Streams

• A stream is a queue of execution
• Non-default streams can be executed in parallel

Time H2D H2D H2D Stream1 Stream2 Stream3 Kernel Kernel Kernel D2H D2H D2H H2D = Host to device transfer D2H = Device to host transfer

Luis Domingues - January 2015 8

Overlapping pixels and SCT

• Use CUDA Streams
• Start the processing of SCT before pixels end

Time Pixel stream SCT stream Kernels Time stamp Time stamp Kernels Time stamp Time stamp

Luis Domingues - January 2015 9

Overlapping pixels and SCT

Luis Domingues - January 2015 10

Overlapping pixels and SCT

• For 2000 events, without overlapping

– Avg Pixel: 2.03 ms – Avg SCT: 1.95 ms – Total avg: 3.98 ms

• For 2000 events, overlapping

– Avg Pixel: 2.3 ms – Avg SCT: 2.5 ms

Luis Domingues - January 2015 11

Overlapping pixels and SCT

• Total execution

– Without overlapping: 8.65 s – With overlapping:

6.53 s

Luis Domingues - January 2015 12

Multi-thread server side

• Huge amount of “small” data

– They do not fulfill the GPU

• Parallelize the “event” level processing with streams

Luis Domingues - January 2015 13

Multi-thread server side

Client Client Client Client Client Client Client Client FIFO

Luis Domingues - January 2015 14

Multi-thread server side

• Life of a thread

Luis Domingues - January 2015 15

Multi-thread server side

Luis Domingues - January 2015 16

Multi-thread server side

• Executions time

– Without overlapping:

8.65 s

– With overlapping:

6.53 s

– Multi-threading server side: 4.7 s

Luis Domingues - January 2015 17

CUDA Occupancy

• A good setup of Grid/Block size in card can be

significant

• CUDA offers an API to maximize the occupancy of the

kernels

Luis Domingues - January 2015 18

CUDA Occupancy

Cuda Core GPU Multiprocessor

Luis Domingues - January 2015 19

CUDA Occupancy

Cuda Core GPU

• Bad block size Setup

Kernel 1 Kernel 2 Intra-block synchronization Multiprocessor

Luis Domingues - January 2015 20

CUDA Occupancy

Cuda Core Multiprocessor GPU

• Better block Setup

Kernel 1 Kernel 2 Intra-block synchronization

Luis Domingues - January 2015 21

CUDA Occupancy

• Maximize the occupancy kills global performances
• Runs results for 2000 events

– Big Blocks size:

10.88 s

– Original configuration:

4.7 s

– Small blocks size:

4.4 s

Luis Domingues - January 2015 22

CUDA Occupancy

• Maximize the occupancy kills global performances
• Runs results for 2000 events

– Big blocks size:

3 kernels in parallel (Max 5)

– Small blocks size:

4 kernels in parallel (Max 7)

Luis Domingues - January 2015 23

Conclusion

• Important points when using a GPU

– Port of an algorithm to the GPU – Communicate with the GPU – Host side design

• Keep the GPU busy
• Big occupancy does not allow the GPU to schedule its

tasks efficiently