incremental run-time application mapping for heterogeneous network on chip 2012 ieee 14th...

50
Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications Jingcheng Shao, Chen Tian-zhou, Li Liu 1

Upload: chad-boyd

Post on 18-Jan-2018

218 views

Category:

Documents


0 download

DESCRIPTION

Outline  Introduction  Near Convex Region Algorithm  Mapping Problem and Evaluation Metrics  Heterogeneous Near Convex Region Algorithm (HNCR)  Experiments and Results  Conclusion 3

TRANSCRIPT

Page 1: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

1

Incremental Run-time Application Mapping for Heterogeneous Network

on Chip2012 IEEE 14th International Conference on High Performance Computing and

CommunicationsJingcheng Shao, Chen Tian-zhou, Li Liu

Page 2: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

2

Outline

Introduction Near Convex Region Algorithm Mapping Problem and Evaluation Metrics Heterogeneous Near Convex Region Algorithm (HNCR) Experiments and Results Conclusion

Page 3: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

3

Outline

Introduction Near Convex Region Algorithm Mapping Problem and Evaluation Metrics Heterogeneous Near Convex Region Algorithm (HNCR) Experiments and Results Conclusion

Page 4: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

4

Introduction

Propose an incremental run-time application mapping algorithm for heterogeneous NoC

Apply the idea of near convex region to heterogeneous NoC

Page 5: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

5

Outline

Introduction Near Convex Region Algorithm Mapping Problem and Evaluation Metrics Heterogeneous Near Convex Region Algorithm (HNCR) Experiments and Results Conclusion

Page 6: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

6

Near Convex Region Algorithm

Two steps Select a near convex region whose area is close to its convex hull Assign nodes to the selected region

Optimizing the mapping results of not only the currently incoming application but also the additional applications in the future

Page 7: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

7

Near Convex Region Algorithm (cont.)

Convex region?

Page 8: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

8

Near Convex Region Algorithm (cont.)

Convex region?

Page 9: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

9

Near Convex Region Algorithm (cont.)

Convex hull

Page 10: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

10

Near Convex Region Algorithm (cont.)

Convex hull

Page 11: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

11

Near Convex Region Algorithm (cont.)

Convex hull

Page 12: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

12

Near Convex Region Algorithm (cont.)

Page 13: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

13

Near Convex Region Algorithm (cont.)

Page 14: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

14

Near Convex Region Algorithm (cont.)

Page 15: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

15

Outline

Introduction Near Convex Region Algorithm Mapping Problem and Evaluation Metrics Heterogeneous Near Convex Region Algorithm (HNCR) Experiments and Results Conclusion

Page 16: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

16

Mapping Problem and Evaluation Metrics

Page 17: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

17

Mapping Problem and Evaluation Metrics Application Communication Graph

ACG = G(V, E) W(ei,j) : communication volume T(vk) : the type of a vertex (Tcpu, Txpu) Wcpu(vk) : computing volume using CPU Wxpu(vk) : computing volume using XPU

Application mapping map(vk) -> PEi,j

MAP(ACG) -> R

Page 18: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

18

Mapping Problem and Evaluation Metrics Energy model

Ecomp : computing energy consumption Ecomm : communication energy consumption

Computing energy Vk is assigned to CPU, then Xk = 1 Vk is assigned to XPU, then Xk = 0

Page 19: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

19

Mapping Problem and Evaluation Metrics Communication energy

Total energy

computing communication

Page 20: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

20

Outline

Introduction Near Convex Region Algorithm Mapping Problem and Evaluation Metrics Heterogeneous Near Convex Region Algorithm (HNCR) Experiments and Results Conclusion

Page 21: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

21

HNCR-Region Selection

Find a proper number of XPU

K = 5*(4/36) Contiguous convex region selection

Page 22: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

22

HNCR-Region Selection D(PE) : the number of

available neighbors of the PE

C(PE) : the distance from the geometric center of the selected region to the PE

Page 23: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

23

HNCR-Region Selection D(PE) : the number of

available neighbors of the PE

C(PE) : the distance from the geometric center of the selected region to the PE

R’

Page 24: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

24

HNCR-Region Selection D(PE) : the number of

available neighbors of the PE

C(PE) : the distance from the geometric center of the selected region to the PE

R’

Page 25: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

25

HNCR-Region Selection D(PE) : the number of

available neighbors of the PE

C(PE) : the distance from the geometric center of the selected region to the PE

R’

Page 26: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

26

HNCR-Region Selection D(PE) : the number of

available neighbors of the PE

C(PE) : the distance from the geometric center of the selected region to the PE

R’

S

Page 27: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

27

HNCR-Region Selection D(PE) : the number of

available neighbors of the PE

C(PE) : the distance from the geometric center of the selected region to the PE

R’

Page 28: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

28

HNCR-Region Selection D(PE) : the number of

available neighbors of the PE

C(PE) : the distance from the geometric center of the selected region to the PE

R’

Page 29: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

29

HNCR-Region Selection D(PE) : the number of

available neighbors of the PE

C(PE) : the distance from the geometric center of the selected region to the PE

R’

SS

Page 30: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

30

HNCR-Region Selection D(PE) : the number of

available neighbors of the PE

C(PE) : the distance from the geometric center of the selected region to the PE

R’

Page 31: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

31

HNCR-Node Allocation Sort the node of application

Step 1 : select all Txpu, sort their computing volume differences in decreasing order

V5, V4 Keep the first K nodes (assume k =1)

Step 2 : sort the remaining nodes by their communication volume with adjacent nodes in decreasing order

V1, V4, V2, V3

Step 3 : append the second list to the tail of the first one V5, V1, V4, V2, V3

Page 32: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

32

HNCR-Node Allocation DISCOVER : Select possible temporary

locations for a node

FINISH : Select an accurate location for a node such that the distance between this node and its “discovered” or “finished” neighbors is minimized

Page 33: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

33

HNCR-Node Allocation DISCOVER : Select possible temporary

locations for a node

FINISH : Select an accurate location for a node such that the distance between this node and its “discovered” or “finished” neighbors is minimized

Page 34: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

34

HNCR-Node Allocation DISCOVER : Select possible temporary

locations for a node

FINISH : Select an accurate location for a node such that the distance between this node and its “discovered” or “finished” neighbors is minimized

Page 35: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

35

HNCR-Node Allocation DISCOVER : Select possible temporary

locations for a node

FINISH : Select an accurate location for a node such that the distance between this node and its “discovered” or “finished” neighbors is minimized

Page 36: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

36

HNCR-Node Allocation DISCOVER : Select possible temporary

locations for a node

FINISH : Select an accurate location for a node such that the distance between this node and its “discovered” or “finished” neighbors is minimized

Page 37: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

37

HNCR-Node Allocation DISCOVER : Select possible temporary

locations for a node

FINISH : Select an accurate location for a node such that the distance between this node and its “discovered” or “finished” neighbors is minimized

Page 38: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

38

HNCR-Node Allocation DISCOVER : Select possible temporary

locations for a node

FINISH : Select an accurate location for a node such that the distance between this node and its “discovered” or “finished” neighbors is minimized

Page 39: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

39

Outline

Introduction Near Convex Region Algorithm Mapping Problem and Evaluation Metrics Heterogeneous Near Convex Region Algorithm (HNCR) Experiments and Results Conclusion

Page 40: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

40

Experiment Setup

Target NoC 6 X 6 mesh

ACG Generation TGFF Vertex : 5-8 Degree of vertex : 1-4

Page 41: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

41

Experiment Setup (cont.)

Comparison algorithm Random Greedy

Simulator Booksim Orion : calculate energy consumption

Page 42: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

42

Experiments and Results

Two performance metrics Average latency Average energy consumption

Page 43: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

43

Injection Rate

Page 44: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

44

Traffic Distributionapplication

Page 45: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

45

Traffic Distribution

Page 46: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

46

Mapping Process

Page 47: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

47

Mapping Process (cont.)

Page 48: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

48

Outline

Introduction Near Convex Region Algorithm Mapping Problem and Evaluation Metrics Heterogeneous Near Convex Region Algorithm (HNCR) Experiments and Results Conclusion

Page 49: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

49

Conclusion

Proposed an incremental run-time application mapping algorithm for heterogeneous NoC

Extend the algorithm to heterogeneous NoC which more types of PEs The algorithm needs to be adjusted when system is much

complicated

Page 50: Incremental Run-time Application Mapping for Heterogeneous Network on Chip 2012 IEEE 14th International Conference on High Performance Computing and Communications

50

Thank you !