tyler sorensen adviser: jade alglave university college london wpli 2015 april 12, 2105 1 gpu...

94
Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

Upload: baldwin-oliver

Post on 15-Jan-2016

215 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

1

Tyler SorensenAdviser: Jade Alglave

University College London

WPLI 2015 April 12, 2105

GPU Concurrency: Weak Behaviours and Programming Assumptions

Page 2: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

2

Based on our ASPLOS ‘15 paper:

Jade Alglave1,2, Mark Batty3, Alastair F. Donaldson4, Ganesh Gopalakrishnan5, Jeroen Ketema4, Daniel Poetzl6, Tyler Sorensen1,5, John Wickerson4

1 University College London, 2 Microsoft Research, 3 University of Cambridge, 4 Imperial College London, 5 University of Utah, 6 University of Oxford

GPU Concurrency: Weak Behaviours and Programming Assumptions

Page 3: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

3

Intel Core i7 4500 CPU

Page 4: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

4

Page 5: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

5

Nvidia Tesla C2075 GPU

Page 6: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

6

Roadmap

• what happened to the pony • how we found the bug • how we are able to fix the pony

(background)(methodology)(contribution)

Page 7: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

7

What happened to the pony?

• the visualization bugs are due to weak memory behaviours on GPUs

Page 8: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

8

Weak memory models

• consider the test known as message passing (mp)• an instance of this test appears in the pony code

Page 9: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

9

Weak memory models

• consider the test known as message passing (mp)• initial state: x and y are memory locations

Page 10: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

10

Weak memory models

• consider the test known as message passing (mp)• thread ids

Page 11: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

11

Weak memory models

• consider the test known as message passing (mp)• program: for each thread id

Page 12: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

12

Weak memory models

• consider the test known as message passing (mp)• assertion: question about the final state of registers

Page 13: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

13

Message passing (mp) test

• Tests how to implement a handshake idiom

Data

Data

Page 14: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

14

Message passing (mp) test

• Tests how to implement a handshake idiom

Flag

Flag

Page 15: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

15

Message passing (mp) test

• Tests how to implement a handshake idiom

Stale Data

Page 16: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

16

Page 17: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

17

Page 18: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

18

Page 19: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

19

Page 20: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

20

Page 21: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

21

assertion cannotbe satisfied by interleavings

this is knownas Lamport’s sequentialconsistency (or SC)

Page 22: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

22

Weak memory models

• can we assume assertion will never pass?

Page 23: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

23

Weak memory models

• can we assume assertion will never pass? No!

Page 24: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

24

Weak memory models

• Alglave and Maranget report this assertion appears 41 million times out of 5 billion test runs on Tegra2 ARM processor1

1http://diy.inria.fr/cats/tables.html

Page 25: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

25

Weak memory models

• what happened?

• architectures implement weak memory models where the hardware is allowed to re-order certain memory instructions.

• weak memory models can allow weak behaviors (executions that do not correspond to an interleaving)

Page 26: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

26

GPU memory models

• what type of memory model do current GPUs implement?

• documentation is sparse

• CUDA has 1 page + 1 example • PTX has 1 page + 0 examples

• given in English prose

• we need to know this if we are to write correct GPU programs!

Page 27: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

27

CTA 0 CTA 1 CTA n

Threads

GPU programming

Global Memory

Shared Memory For CTA 0

Shared Memory For CTA 1

Shared Memory For CTA n

Within CTAs, threadsare grouped into warps(32 threads per warp in Nvidia GPUs)

Page 28: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

28

Threads

GPU programming

Global Memory

Page 29: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

29

CTA 0 CTA 1 CTA n

Threads

GPU programming

Global Memory

Page 30: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

30

CTA 0 CTA 1 CTA n

Threads

GPU programming

Global Memory

Shared Memory For CTA 0

Shared Memory For CTA 1

Shared Memory For CTA n

Page 31: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

31

CTA 0 CTA 1 CTA n

Threads

GPU programming

Global Memory

Shared Memory For CTA 0

Shared Memory For CTA 1

Shared Memory For CTA n

Within CTAs, threadsare grouped into warps(32 threads per warp in Nvidia GPUs)

Page 32: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

32

(background)(methodology)(contribution)

Roadmap

• what happened to the pony • how we found the bug • how we are able to fix the pony

Page 33: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

33

Methodology

GPU litmus tests

GPU hardware

formal model

compare results

Page 34: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

34

GPU tests

• GPU litmus test considerations

Scope Tree (device (cta T0) (cta T1) )x: global, y: global

Page 35: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

35

GPU tests

• GPU litmus test considerations• PTX instructions

Scope Tree (device (cta T0) (cta T1) )x: global, y: global

Page 36: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

36

GPU tests

• GPU litmus test considerations• what memory region (shared or global) are x and y in?

Scope Tree (device (cta T0) (cta T1) )x: global, y: global

Page 37: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

37

GPU tests

• GPU litmus test considerations• what memory region (shared or global) are x and y in?

Page 38: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

38

GPU tests

• GPU litmus test considerations• are T0 and T1 in the same CTA or different CTAs?

Scope Tree (device (cta T0) (cta T1) )x: global, y: global

Page 39: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

39

GPU tests

• GPU litmus test considerations• are T0 and T1 in the same CTA or different CTAs?

Page 40: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

40

Running tests

• we extend the litmus CPU testing tool of Alglave and Maranget to run GPU tests

• given a GPU litmus test, generates an executable CUDA or OpenCL code for the test

Page 41: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

41

Heuristics

• memory stress: extra threads read and write to scratch memory

T0 T1 extra thread 1 extra thread n . . . . .

run T0 test program

run T1 test program

loop:read or write to scratchpad

loop:read or write to scratchpad

Page 42: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

42

Heuristics

• random threads: randomize the location of threads

T0

T1

Page 43: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

43

Heuristics

• random threads: randomize the location of threads

Page 44: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

44

Heuristics

• random threads: randomize the location of threads

Page 45: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

45

Heuristics

• random threads: randomize the location of threads

Page 46: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

46

Heuristics

test none random threads memory stress

memory stress +

random threads

gpu-mp 0

# of weak behaviours in 100,000 runs for different heuristics on a Nvidia Tesla C2075

Page 47: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

47

Heuristics

test none random threads memory stress

memory stress +

random threads

gpu-mp 0 0

# of weak behaviours in 100,000 runs for different heuristics on a Nvidia Tesla C2075

Page 48: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

48

Heuristics

test none random threads memory stress

memory stress +

random threads

gpu-mp 0 0 139

# of weak behaviours in 100,000 runs for different heuristics on a Nvidia Tesla C2075

Page 49: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

49

Heuristics

test none random threads memory stress

memory stress +

random threads

gpu-mp 0 0 139 522

# of weak behaviours in 100,000 runs for different heuristics on a Nvidia Tesla C2075

Page 50: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

50

How we found the pony bug

test none random threads memory stress

memory stress +

random threads

gpu-mp 0 0 139 522

This is the idiom and heuristics that caused bug!

Page 51: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

51

(background)(methodology)(contribution)

Roadmap

• what happened to the pony• how we found the bug • how we are able to fix the pony

Page 52: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

52

GPU fences

• PTX gives 2 fences to disallow reading stale data

• membar.cta – gives ordering intra-CTA

• membar.gl – gives ordering over device

Page 53: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

53

GPU fences

• Test amended with a parameterizable fence

Scope Tree (device (cta T0) (cta T1) )x: global, y: global

Page 54: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

54

GPU fences

test none membar.cta membar.gl

gpu-mp 3380

# of weak behaviours in 100,000 runs for different fences on a Nvidia Tesla C2075

Page 55: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

55

GPU fences

test none membar.cta membar.gl

gpu-mp 3380 2

# of weak behaviours in 100,000 runs for different fences on a Nvidia Tesla C2075

Page 56: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

56

GPU fences

test none membar.cta membar.gl

gpu-mp 3380 2 0

# of weak behaviours in 100,000 runs for different fences on a Nvidia Tesla C2075

Page 57: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

57

How do we fix the pony

Tesla C2075 Nvidia GPU

Page 58: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

58

How do we fix the pony

• adding fences to the code

Tesla C2075 Nvidia GPU(with fences)

Page 59: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

59

GPU testing campaign

• we extend the diy CPU litmus test generation tool of Alglave and Maranget to generate GPU tests

• generates litmus tests based on cycles

• enumerates the tests over the GPU thread and memory hierarchy

Page 60: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

60

GPU testing campaign

• Using our tools, we generated and ran 10930 tests over 5 Nvidia chips:

chip year architecture

GTX 750 ti 2014 Maxwell

GTX Titan 2013 Kepler

GTX 660 2012 Kepler

GTX 540m 2011 Fermi

Tesla C2075 2011 Fermi

Page 61: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

61

GPU testing campaign

• Results are hosted at:http://virginia.cs.ucl.ac.uk/sunflowers/asplos15/flat.html

Page 62: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

62

Modeling

• we extended the CPU axiomaitic memory modeling toolherd of Alglave and Maranget, for GPUs

• we developed an axiomatic memory model for PTX which is able to simulate all of our tests

• our model is sound with respect to all of our hardware observations

Page 63: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

63

Modeling

• Demo of web interface

Page 64: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

64

More results

• surprising and buggy behaviours observed:

• GPU mutex implementations allow stale data to be read(found in CUDA by Example book and other academic papers1,2)

led to an erratum issued by Nvidia

• Hardware re-orders loads from the same address in Nvidia Fermi and Kepler

• Some testing on AMD GPUs

1J. A. Stuart and J. D. Owens, "Efficient synchronization primitives for GPUs" CoRR, 2011, http://arxiv.org/pdf/1110.4623.pdf.2B. He and J. X. Yu, “High-throughput transaction executions on graphics processors” PVLDB 2011.

Page 65: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

65

Related work (CPU memory models)• Alglave et. al. have done extensive work on testing and modeling

CPUs (notably IBM Power and ARM) and create the tools diy, litmus, and herd which we extended for this work

• Collier tested CPU memory models using the ARCHTEST tool

Page 66: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

66

Related work (GPU memory models)• Hower et. al. have proposed several SC for race-free language level

memory models for GPUs

Page 67: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

Questions?

Nvidia Tesla C2075 GPU(with fences)

Nvidia Tesla C2075 GPUIntel Core i7 4500 CPU

project page: http://virginia.cs.ucl.ac.uk/sunflowers/asplos15/

Page 68: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

68

CUDA by Example

Intel Core i7 4500 CPU

Page 69: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

69

CUDA by Example

Nvidia Tesla C2075 GPU

Page 70: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

70

CUDA by Example

Nvidia Tesla C2075 GPU(with fences)

Page 71: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

71

Read-after-Read Hazard

Page 72: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

72

Ignore after this

Page 73: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

73

Results

• Surprising and buggy behaviours observed:

• SC-per-location violations on NVIDIA Fermi and Kepler architecture:

todo: add CORR test

Page 74: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

74

Limitations

• warps: we do not test intra-warp behaviours as the lock step behaviour of warps is not compatible with some of our heuristics

• grids: we do not test inter-grid behaviours as we did not find any examples in the literature

Page 75: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

75

GPU programming

• GPUs are SIMT (Single Instruction, Multiple Thread)

• Nvidia GPUs may be programmed using CUDA or OpenCL

Page 76: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

76

Roadmap

• background and motivation• approach• GPU tests• running tests• modeling

Page 77: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

77

Heuristics

• two additional heuristics:

• synchronization: testing threads synchronize immediately before running the test program

• general bank conflicts: generate memory access that conflict with the accesses in the memory stress heuristic

Page 78: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

78

Challenges

• PTX optimizing assembler may reorder or remove instructions

• We developed a tool optcheck which compares the litmus test with the binary and checks for optimizations

Page 79: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

79

Roadmap

• background and motivation• approach• GPU tests• running tests• modeling

Page 80: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

80

GPU tests

• concrete GPU test

T0 | T1 ;

st.cg.s32 [x], 1 | ld.cg.s32 r1,[y] ;

st.cg.s32 [y], 1 | ld.cg.s32 r2,[x] ;

ScopeTree

(grid(cta(warp T0) (warp T1)))

x: shared, y: global

exists (1:r1=1 /\ 1:r2=0)

Page 81: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

81

GPU tests

• concrete GPU test

T0 | T1 ;st.cg.s32 [x], 1 | ld.cg.s32 r1,[y] ;st.cg.s32 [y], 1 | ld.cg.s32 r2,[x] ;

ScopeTree(grid(cta(warp T0) (warp T1)))

x: shared, y: global

exists (1:r1=1 /\ 1:r2=0)

Page 82: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

82

GPU tests

• concrete GPU test

T0 | T1 ;st.cg.s32 [x], 1 | ld.cg.s32 r1,[y] ;st.cg.s32 [y], 1 | ld.cg.s32 r2,[x] ;

ScopeTree(grid(cta(warp T0) (warp T1)))

x: shared, y: global

exists (1:r1=1 /\ 1:r2=0)

Page 83: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

83

GPU programming

explicit hierarchical concurrency model

• thread hierarchy:• thread

• warp

• CTA (Cooperative Thread Array)

• grid

• memory hierarchy:• shared memory

• global memory

Page 84: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

84

GPU background

Images from Wikipedia [15,16,17]

• GPU is a highly parallel co-processor

• currently found in devicesfrom tablets to top supercomputers

• not just used for visualization anymore!

Page 85: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

85

References

[1] L. Lamport, "How to make a multiprocessor computer that correctly executes multi-process programs" Trans. Comput. 1979.

[2] J. Alglave, L. Maranget, S. Sarkar, and P. Sewell, "Litmus: Running tests against hardware" TACAS 2011.

[3] J. Alglave, L. Maranget, and M. Tautschnig, "Herding cats: modelling, simulation, testing, and data-mining for weak memory" TOPLAS 2014.

[4] NVIDIA, "CUDA C programming guide, version 6 (July 2014)" http://docs.nvidia.com/cuda/pdf/CUDA C Programming Guide.pdf

[5] NVIDIA, "Parallel Thread Execution ISA: Version 4.0 (Feb. 2014)," http://docs.nvidia.com/cuda/parallel-thread-execution

[6] J. Alglave, L. Maranget, S. Sarkar, and P. Sewell, “Fences in weak memory models (extended version)” FMSD 2012

[7] J. Sanders and E. Kandrot, “CUDA by Example: An Introduction to General-Purpose GPU Programming” Addison-Wesley Professional, 2010.

Page 86: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

86

References

[8] J. A. Stuart and J. D. Owens, "Efficient synchronization primitives for GPUs" CoRR, 2011, http://arxiv.org/pdf/1110.4623.pdf.

[9] B. He and J. X. Yu, “High-throughput transaction executions on graphics processors” PVLDB 2011.

[10] W. W. Collier, Reasoning About Parallel Architectures. Prentice-Hall, Inc., 1992.

[11] D. R. Hower, B. M. Beckmann, B. R. Gaster, B. A. Hechtman, M. D. Hill, S. K. Reinhardt, and D. A. Wood, "Sequential consistency for heterogeneous-race-free" MSPC 2013.

[12] D. R. Hower, B. A. Hechtman, B. M. Beckmann, B. R. Gaster, M. D. Hill, S. K. Reinhardt, and D. A. Wood, "Heterogeneous-race-free memory models," ASPLOS 2014

[13] T. Sorensen, G. Gopalakrishnan, and V. Grover, "Towards shared memory consistency models for GPUs" ICS 2013

[14] W.-m. W. Hwu, “GPU Computing Gems Jade Edition” Morgan Kaufmann Publishers Inc., 2011.

Page 87: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

87

References

[15] http://en.wikipedia.org/wiki/Samsung_Galaxy_S5

[16] http://en.wikipedia.org/wiki/Titan_(supercomputer)

[17] http://en.wikipedia.org/wiki/Barnes_Hut_simulation

Page 88: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

88

Roadmap

• what happened to the pony (background)• how we found the bug (methodology)• how we are able to fix the pony (contribution)

Page 89: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

89

Message passing (mp) test

• Tests how to implement a handshake idiom• Found in Octree code for the pony visualization

Page 90: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

90

Message passing (mp) test

• Tests how to implement a handshake idiom

Data

Data

Page 91: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

91

Message passing (mp) test

• Tests how to implement a handshake idiom

Flag

Flag

Page 92: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

92

Methodology

• empirically explore the hardware memory model implemented on deployed NVIDIA and AMD GPUs

• develop hardware memory model testing tools for GPUs

• analyze classic (i.e. CPU) memory model properties and communication idioms in CUDA applications

• run large families of tests on GPUs as a basis for modeling and bug hunting

Page 93: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

93

Message passing (mp) test

• Tests how to implement a handshake idiom

Stale Data

Page 94: Tyler Sorensen Adviser: Jade Alglave University College London WPLI 2015 April 12, 2105 1 GPU Concurrency: Weak Behaviours and Programming Assumptions

94

Running tests

• however, unlike CPUs, simply running the tests did not yield any weak memory behaviours for Nvidia chips!

• we developed heuristics to run tests under a variety of stress to expose weak behaviours