A Runtime System for Generalized Committed Choice

Presentation @ APPLC'12 (affiliated with PLDI'12)

Transcript

1. A Runtime System for Generalized Committed Choice Xiao Jia Kenny Q. Zhu Shanghai Jiao Tong University {xjia,kzhu}@cs.sjtu.edu.cn Joxan Jaffar Roland H.C. Yap National University of Singapore {joxan,ryap}@comp.nus.edu.sg
2. /53 Outline •  Motivation •  Generalized committed choice •  Optimization: multiple universe •  Evaluation June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 2
3. /53 Outline •  Motivation •  Generalized committed choice •  Optimization: multiple universe •  Evaluation June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 3
4. /53 Motivation •  Online trading system •  Bob wants to upgrade his camera •  Jill wants to downgrade June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 4
5. /53 •  Bob: (upgrade) 1.  buy ( good lens ) ; sell ( average lens ) 2.  buy ( good cam ) ; sell ( average cam ) •  Jill: (downgrade) 1.  sell ( good lens ) ; buy ( average cam ) 2.  sell ( good cam ) ; buy ( average cam ) June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 5
6. /53 •  Bob: (upgrade) 1.  buy ( good lens ) ; sell ( average lens ) 2.  buy ( good cam ) ; sell ( average cam ) •  Jill: (downgrade) 1.  sell ( good lens ) ; buy ( average cam ) 2.  sell ( good cam ) ; buy ( average cam ) June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 6
7. /53 •  Bob: (upgrade) 1.  buy ( good lens ) ; sell ( average lens ) 2.  buy ( good cam ) ; sell ( average cam ) •  Jill: (downgrade) 1.  sell ( good lens ) ; buy ( average cam ) 2.  sell ( good cam ) ; buy ( average cam ) June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 7
8. /53 •  Bob: (upgrade) 1.  buy ( good lens ) ; sell ( average lens ) 2.  buy ( good cam ) ; sell ( average cam ) •  Jill: (downgrade) 1.  sell ( good lens ) ; buy ( average cam ) 2.  sell ( good cam ) ; buy ( average cam ) June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 8
9. /53 •  Bob: (upgrade) 1.  buy ( good lens ) ; sell ( average lens ) 2.  buy ( good cam ) ; sell ( average cam ) •  Jill: (downgrade) 1.  sell ( good lens ) ; buy ( average cam ) 2.  sell ( good cam ) ; buy ( average cam ) June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 9
10. /53 •  Bob: (upgrade) 1.  buy ( good lens ) ; sell ( average lens ) 2.  buy ( good cam ) ; sell ( average cam ) •  Jill: (downgrade) 1.  sell ( good lens ) ; buy ( average cam ) 2.  sell ( good cam ) ; buy ( average cam ) June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 10
11. /53 The transaction can be completed only if both Bob and Jill choose the second choice June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 11
12. /53 However, it is impossible for the agents to know the others’ choices in the context of online trading June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 12
13. /53 So you have no good choice ! June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 13
14. /53 June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 14 But we want maximal happiness among the agents
15. /53 June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 15 What if we can try different combinations of the choices in isolated environments?
16. /53 June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 16 Bob: 1. buy ( good lens ) sell ( average lens ) Jill: 1. sell ( good lens ) buy ( average cam ) Bob: 1. buy ( good lens ) sell ( average lens ) Jill: 2. sell ( good cam ) buy ( average cam ) Bob: 2. buy ( good cam ) sell ( average cam ) Jill: 1. sell ( good lens ) buy ( average cam ) Bob: 2. buy ( good cam ) sell ( average cam ) Jill: 2. sell ( good cam ) buy ( average cam )
17. /53 June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 17 Bob: 1. buy ( good lens ) sell ( average lens ) Jill: 1. sell ( good lens ) buy ( average cam ) Bob: 1. buy ( good lens ) sell ( average lens ) Jill: 2. sell ( good cam ) buy ( average cam ) Bob: 2. buy ( good cam ) sell ( average cam ) Jill: 1. sell ( good lens ) buy ( average cam ) Bob: 2. buy ( good cam ) sell ( average cam ) Jill: 2. sell ( good cam ) buy ( average cam ) 4 speculative worlds
18. /53 June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 18 Bob: 1. buy ( good lens ) sell ( average lens ) Jill: 1. sell ( good lens ) buy ( average cam ) Bob: 1. buy ( good lens ) sell ( average lens ) Jill: 2. sell ( good cam ) buy ( average cam ) Bob: 2. buy ( good cam ) sell ( average cam ) Jill: 1. sell ( good lens ) buy ( average cam ) Bob: 2. buy ( good cam ) sell ( average cam ) Jill: 2. sell ( good cam ) buy ( average cam ) REAL!  
19. /53 Write agent programs like this June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 19 // bob.cc int main() { gcc(lens, cam); return 0; } void lens() { buy(GoodLens); sell(AvgLens); cm(); } void cam() { buy(GoodCam); sell(AvgCam); cm(); } // jill.cc int main() { gcc(lens, cam); return 0; } void lens() { sell(GoodLens); buy(AvgCam); cm(); } void cam() { buy(GoodCam); sell(AvgCam); cm(); } buy and sell are blocking and will be implemented in a few slides later
20. /53 Agent programs run like this June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 20 Bob Bob: lens() Jill Bob: lens() Jill: lens() Bob: lens() Jill: cam() Bob: cam() Jill Bob: cam() Jill: lens() Bob: cam() Jill: cam() w1 w2 w3 w4 (Suppose Bob starts first)
21. /53 Agent programs run like this June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 21 Bob Bob: lens() Jill Bob: lens() Jill: lens() Bob: lens() Jill: cam() Bob: cam() Jill Bob: cam() Jill: lens() Bob: cam() Jill: cam() w1 w2 w3 w4 (Suppose Bob starts first)
22. /53 Outline •  Motivation •  Generalized committed choice •  Optimization: multiple universe •  Evaluation June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 22
23. /53 Programming Model = Coordination Model + Computation Model June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 23 David Gelernter, Nicholas Carriero. Coordination languages and their significance. 1992.
24. /53 Programming Model = Coordination Model + Computation Model June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 24 David Gelernter, Nicholas Carriero. Coordination languages and their significance. 1992. Traditional computations like additions and multiplications …
25. /53 Programming Model = Coordination Model + Computation Model June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 25 David Gelernter, Nicholas Carriero. Coordination languages and their significance. 1992. For coordination of computations Sequential control flows (if/for/…) Or concurrent coordination model for multiple agents
26. /53 Generalized Committed Choice •  A concurrent coordination model •  Speculation among concurrent agents by using multi-worlds •  Worlds: isolated tuple spaces June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 26
27. /53 Generalized Committed Choice •  A concurrent coordination model •  Speculation among concurrent agents by using multi-worlds •  Worlds: isolated tuple spaces June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 27 (tree  structure)  
28. /53 Generalized Committed Choice •  A concurrent coordination model •  Speculation among concurrent agents by using multi-worlds •  Worlds: isolated tuple spaces June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 28 (Sell, GoodLens, 123) (Sell, GoodLens, 456) (Ack, GoodCam, 789) (a  collec2on  of  tuples)   ……  
29. /53 The Runtime System for GCC •  A central server for coordination •  Agent programs act as clients •  Clients communicate with the server via network connections June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 29
30. /53 GCC Primitives gcc fork multiple choices cm commit me (commit to this choice) cu commit you (kill this choice) in reads and removes a tuple from a tuple space rd non-destructively reads a tuple space out produces a tuple, writing it into a tuple space June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 30
31. /53 Write agent programs like this June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 31 // bob.cc int main() { gcc(lens, cam); return 0; } void lens() { buy(GoodLens); sell(AvgLens); cm(); } void cam() { buy(GoodCam); sell(AvgCam); cm(); } // jill.cc int main() { gcc(lens, cam); return 0; } void lens() { sell(GoodLens); buy(AvgCam); cm(); } void cam() { buy(GoodCam); sell(AvgCam); cm(); }
32. /53 Example: Implement sell/buy void sell(Product t) { int id = get_unique_id(); out("Sell %d %d", t, id); in ("Buy _ %d", id); out("Ack %d %d", t, id); } void buy(Product t) { int id = -1; in ("Sell %d &d", t, &id); out("Buy %d %d", t, id); rd ("Ack _ %d", id); } June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 32 enum Product { GoodLens, AvgLens, GoodCam, AvgCam };
33. /53 Example: Implement sell/buy void sell(Product t) { int id = get_unique_id(); out("Sell %d %d", t, id); in ("Buy _ %d", id); out("Ack %d %d", t, id); } void buy(Product t) { int id = -1; in ("Sell %d &d", t, &id); out("Buy %d %d", t, id); rd ("Ack _ %d", id); } June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 33 enum Product { GoodLens, AvgLens, GoodCam, AvgCam }; in/rd/out are primitives for tuple operations
34. /53 Example: Implement sell/buy void sell(Product t) { int id = get_unique_id(); out("Sell %d %d", t, id); in ("Buy _ %d", id); out("Ack %d %d", t, id); } void buy(Product t) { int id = -1; in ("Sell %d &d", t, &id); out("Buy %d %d", t, id); rd ("Ack _ %d", id); } June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 34 enum Product { GoodLens, AvgLens, GoodCam, AvgCam }; Tuple
35. /53 Example: Implement sell/buy void sell(Product t) { int id = get_unique_id(); out("Sell %d %d", t, id); in ("Buy _ %d", id); out("Ack %d %d", t, id); } void buy(Product t) { int id = -1; in ("Sell %d &d", t, &id); out("Buy %d %d", t, id); rd ("Ack _ %d", id); } June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 35 enum Product { GoodLens, AvgLens, GoodCam, AvgCam }; Tuples are in the form of (operation, product, trade id)
36. /53 Example: Implement sell/buy void sell(Product t) { int id = get_unique_id(); out("Sell %d %d", t, id); in ("Buy _ %d", id); out("Ack %d %d", t, id); } void buy(Product t) { int id = -1; in ("Sell %d &d", t, &id); out("Buy %d %d", t, id); rd ("Ack _ %d", id); } June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 36 enum Product { GoodLens, AvgLens, GoodCam, AvgCam }; Underscore ( _ ) matches any value Ampersand ( & ) returns a value
37. /53 Example: Implement sell/buy void sell(Product t) { int id = get_unique_id(); out("Sell %d %d", t, id); in ("Buy _ %d", id); out("Ack %d %d", t, id); } void buy(Product t) { int id = -1; in ("Sell %d &d", t, &id); out("Buy %d %d", t, id); rd ("Ack _ %d", id); } June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 37 enum Product { GoodLens, AvgLens, GoodCam, AvgCam }; Underscore ( _ ) matches any value Ampersand ( & ) returns a value Also  support  range  queries  on   tuple  elements  (e.g.  prices)   See  details  in  paper  J  
38. /53 Outline •  Motivation •  Generalized committed choice •  Optimization: multiple universe •  Evaluation June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 38
39. /53 Agent programs run like this June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 39 Bob Bob: lens() Jill Bob: lens() Jill: lens() Bob: lens() Jill: cam() Bob: cam() Jill Bob: cam() Jill: lens() Bob: cam() Jill: cam() w1 w2 w3 w4
40. /53 Exponential ! June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 40 For 10 instances each of Bob and Jill agents, the theoretical bound is 220 worlds
41. /53 Exponential ! June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 41 If we can process 100 worlds per second, it will take 3 hours for only 20 agents
42. /53 Not all the agents are trading cameras and lenses Other agents may trade apples and bananas June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 42
43. /53 Without multiple universe June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 43 ……   ……  
44. /53 With multiple universe June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 44 ……   ……   Camera  Universe   Fruits  Universe  
45. /53 With multiple universe June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 45 ……   ……   220 >> 210 + 210 Details in paper J Camera  Universe   Fruits  Universe  
46. /53 Outline •  Motivation •  Generalized committed choice •  Optimization: multiple universe •  Evaluation June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 46
47. /53 10 instances of each Bob and Jill agents June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 47 Bob-Jill cases min max Max # of worlds 11 67 Max storage (MB) 40.6 53.5 Response time (s) maximum 9.22 133.62 average 4.44 44.48 Turnaround time (s) maximum 9.23 1011.91 average 5.07 318.19
48. /53 10 instances of each Bob and Jill agents June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 48 Bob-Jill cases min max Max # of worlds 11 67 Max storage (MB) 40.6 53.5 Response time (s) maximum 9.22 133.62 average 4.44 44.48 Turnaround time (s) maximum 9.23 1011.91 average 5.07 318.19 Worst-­‐case   Best-­‐case  
49. /53 10 instances of each Bob and Jill agents June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 49 Bob-Jill cases min max Max # of worlds 11 67 Max storage (MB) 40.6 53.5 Response time (s) maximum 9.22 133.62 average 4.44 44.48 Turnaround time (s) maximum 9.23 1011.91 average 5.07 318.19 Much  smaller  than  220  
50. /53 10 instances of each Bob and Jill agents June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 50 Bob-Jill cases min max Max # of worlds 11 67 Max storage (MB) 40.6 53.5 Response time (s) maximum 9.22 133.62 average 4.44 44.48 Turnaround time (s) maximum 9.23 1011.91 average 5.07 318.19 Caused  by  the  opportunisEc   forking  opEmizaEon   A  tradeoff  between  usage  of   system  resources  versus  latency  
51. /53 Conclusion •  Provide more possibilities •  Practical and realistic •  Independent and embeddable in mainstream languages (C/C++/…) June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 51
52. /53 Demo June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 52
53. /53 Thank you for listening ! June 15, 2012 Xiao Jia <xjia@cs.sjtu.edu.cn> 53 Contact: Xiao Jia (贾枭) xjia@cs.sjtu.edu.cn