Bridging the GAP: Towards Approximate Graph Analytics

Transcript

1. Bridging the GAP: Towards Approximate Graph Analytics Anand Iyer ⋆, Aurojit Panda ▪, Shivaram Venkataraman⬩, Mosharaf Chowdhury▴, Aditya Akella⬩, Scott Shenker ⋆, Ion Stoica ⋆ ⋆ UC Berkeley ▪ NYU ⬩ University of Wisconsin ▴ University of Michigan June 10, 2018
2. Graphs popular in big data analytics
3. Graphs popular in big data analytics
4. Graphs popular in big data analytics
5. Graphs popular in big data analytics
7. Can benefit from timely analysis
8. Can benefit from timely analysis Often do not require exact answers
9. Cellular Network Analytics
10. Financial Network Analytics Image courtesy: Neo4J
11. Graph Analytics
12. Graph Analytics Takes several minutes to produce exact answers
13. Can we leverage approximation for graph analytics?
14. Apply query on samples of the input data Approximate Analytics
15. Apply query on samples of the input data Approximate Analytics ID City Buff Ratio 1 NYC 0.78 2 NYC 0.13 3 Berkeley 0.25 4 NYC 0.19 5 NYC 0.11 6 Berkeley 0.09 7 NYC 0.18 8 NYC 0.15 9 Berkeley 0.13 10 Berkeley 0.49 11 NYC 0.19 12 Berkeley 0.10
16. Apply query on samples of the input data Approximate Analytics ID City Buff Ratio 1 NYC 0.78 2 NYC 0.13 3 Berkeley 0.25 4 NYC 0.19 5 NYC 0.11 6 Berkeley 0.09 7 NYC 0.18 8 NYC 0.15 9 Berkeley 0.13 10 Berkeley 0.49 11 NYC 0.19 12 Berkeley 0.10 What is the average buffering ratio in the table?
17. Apply query on samples of the input data Approximate Analytics ID City Buff Ratio 1 NYC 0.78 2 NYC 0.13 3 Berkeley 0.25 4 NYC 0.19 5 NYC 0.11 6 Berkeley 0.09 7 NYC 0.18 8 NYC 0.15 9 Berkeley 0.13 10 Berkeley 0.49 11 NYC 0.19 12 Berkeley 0.10 What is the average buffering ratio in the table? 0.2325
18. Apply query on samples of the input data Approximate Analytics ID City Buff Ratio 1 NYC 0.78 2 NYC 0.13 3 Berkeley 0.25 4 NYC 0.19 5 NYC 0.11 6 Berkeley 0.09 7 NYC 0.18 8 NYC 0.15 9 Berkeley 0.13 10 Berkeley 0.49 11 NYC 0.19 12 Berkeley 0.10 What is the average buffering ratio in the table? ID City Buff Ratio Sampling Rate 2 NYC 0.13 1/4 6 Berkeley 0.25 1/4 8 NYC 0.19 1/4 Uniform Sample
19. Apply query on samples of the input data Approximate Analytics ID City Buff Ratio 1 NYC 0.78 2 NYC 0.13 3 Berkeley 0.25 4 NYC 0.19 5 NYC 0.11 6 Berkeley 0.09 7 NYC 0.18 8 NYC 0.15 9 Berkeley 0.13 10 Berkeley 0.49 11 NYC 0.19 12 Berkeley 0.10 What is the average buffering ratio in the table? ID City Buff Ratio Sampling Rate 2 NYC 0.13 1/4 6 Berkeley 0.25 1/4 8 NYC 0.19 1/4 Uniform Sample 0.19 0.2325
20. Apply query on samples of the input data Approximate Analytics ID City Buff Ratio 1 NYC 0.78 2 NYC 0.13 3 Berkeley 0.25 4 NYC 0.19 5 NYC 0.11 6 Berkeley 0.09 7 NYC 0.18 8 NYC 0.15 9 Berkeley 0.13 10 Berkeley 0.49 11 NYC 0.19 12 Berkeley 0.10 What is the average buffering ratio in the table? ID City Buff Ratio Sampling Rate 2 NYC 0.13 1/4 6 Berkeley 0.25 1/4 8 NYC 0.19 1/4 Uniform Sample 0.19 0.2325 +/- 0.05
21. Apply query on samples of the input data Approximate Analytics ID City Buff Ratio 1 NYC 0.78 2 NYC 0.13 3 Berkeley 0.25 4 NYC 0.19 5 NYC 0.11 6 Berkeley 0.09 7 NYC 0.18 8 NYC 0.15 9 Berkeley 0.13 10 Berkeley 0.49 11 NYC 0.19 12 Berkeley 0.10 What is the average buffering ratio in the table? ID City Buff Ratio Sampling Rate 2 NYC 0.13 1/2 3 Berkeley 0.25 1/2 5 NYC 0.11 1/2 6 Berkeley 0.09 1/2 8 NYC 0.15 1/2 12 Berkeley 0.10 1/2 Uniform Sample
22. Apply query on samples of the input data Approximate Analytics ID City Buff Ratio 1 NYC 0.78 2 NYC 0.13 3 Berkeley 0.25 4 NYC 0.19 5 NYC 0.11 6 Berkeley 0.09 7 NYC 0.18 8 NYC 0.15 9 Berkeley 0.13 10 Berkeley 0.49 11 NYC 0.19 12 Berkeley 0.10 What is the average buffering ratio in the table? ID City Buff Ratio Sampling Rate 2 NYC 0.13 1/2 3 Berkeley 0.25 1/2 5 NYC 0.11 1/2 6 Berkeley 0.09 1/2 8 NYC 0.15 1/2 12 Berkeley 0.10 1/2 Uniform Sample 0.19 +/- 0.05 0.2325 0.22 +/- 0.02
23. Can we use the same idea on graphs? Approximate Analytics on Graphs
24. Can we use the same idea on graphs? Approximate Analytics on Graphs 0 1 4 2 3 graph
25. Can we use the same idea on graphs? Approximate Analytics on Graphs 0 1 4 2 3 edge sampling (p=0.5) graph 0 1 4 2 3
26. Can we use the same idea on graphs? Approximate Analytics on Graphs 0 1 4 2 3 edge sampling (p=0.5) graph e = 1 0 1 4 2 3 triangle counting
27. Can we use the same idea on graphs? Approximate Analytics on Graphs 0 1 4 2 3 edge sampling (p=0.5) graph e = 1 0 1 4 2 3 triangle counting result $ % 2 = 2
28. Answer: 10 Can we use the same idea on graphs? Approximate Analytics on Graphs 0 1 4 2 3 edge sampling (p=0.5) graph e = 1 0 1 4 2 3 triangle counting result $ % 2 = 2
29. Challenge: Non-linear relation between sample size and runtime / error Approximate Analytics on Graphs
30. Challenge: Non-linear relation between sample size and runtime / error Approximate Analytics on Graphs ��� ��� ��� ��� ��� ��� ��� ��� � �� �� �� �� �� �� �� �� �� ������� ����� ������� ���
31. Challenge: Non-linear relation between sample size and runtime / error Approximate Analytics on Graphs How to sample graphs? What is the right sample size? How to compute the error for a given (iterative) graph query?
32. Our Proposal: GAP Run A within T sec Result, Error Graph Algorithms Sparsification Selector Models Sparsifier
33. Our Proposal: GAP Run A within T sec Result, Error Graph Algorithms Sparsification Selector Models Sparsifier
34. Our Proposal: GAP Run A within T sec Result, Error Graph Algorithms Sparsification Selector Models Sparsifier
35. Our Proposal: GAP Run A within T sec Result, Error Graph Algorithms Sparsification Selector Models Sparsifier
36. * Daniel A. Spielman and Shang-Hua Teng. “Spectral Sparsification of Graphs” Sampling for Graph Approximation § Sparsification extensively studied in graph theory § Idea: approximate the graph using a sparse, much smaller graph § Many computationally intensive § Not amenable to distributed implementation § Build on Spielman & Teng’s work* § Keep edges with probability cial properties of the input graph. While several proposals on the type of sparsi￿er exists, many o m are either computationally intensive, or are not amenable to stributed implementation (which is the focus of our work)3. A nitial solution, we developed a simple sparsi￿er adapted from work of Spielman and Teng [31] that is based on vertex degree sparsi￿er uses the following probability to decide to keep an e between vertex a and b: dAV G ⇥ s min(d o a,d i b ) (1 o
37. * Daniel A. Spielman and Shang-Hua Teng. “Spectral Sparsification of Graphs” Sampling for Graph Approximation § Sparsification extensively studied in graph theory § Idea: approximate the graph using a sparse, much smaller graph § Many computationally intensive § Not amenable to distributed implementation § Build on Spielman & Teng’s work* § Keep edges with probability cial properties of the input graph. While several proposals on the type of sparsi￿er exists, many o m are either computationally intensive, or are not amenable to stributed implementation (which is the focus of our work)3. A nitial solution, we developed a simple sparsi￿er adapted from work of Spielman and Teng [31] that is based on vertex degree sparsi￿er uses the following probability to decide to keep an e between vertex a and b: dAV G ⇥ s min(d o a,d i b ) (1 o Exploring other sparsification techniques
38. Estimating the Error / Latency What is the error / speedup due to sparsification? Many approaches in approximate processing literature: • Exhaustively run every possible point • Theoretical closed-bound solutions • Experiment design / Bayesian techniques
39. Estimating the Error / Latency What is the error / speedup due to sparsification? Many approaches in approximate processing literature: • Exhaustively run every possible point • Theoretical closed-bound solutions • Experiment design / Bayesian techniques None applicable for graph approximation
40. Building a model for s Use machine learning to build a model.
41. Building a model for s Model Builder Use machine learning to build a model.
42. Building a model for s Model Builder Use machine learning to build a model. Learn the relation between s and error / latency
43. Building a model for s Model Builder Input features Use machine learning to build a model. Learn the relation between s and error / latency
44. Building a model for s Model Builder Input features Model Use machine learning to build a model. Learn the relation between s and error / latency
45. Building a model for s Model Builder Input features Model Use machine learning to build a model. Learn the relation between s and error / latency “The most important determinant of graph workload characteristics is typically the input graph and surprisingly not the implementation or even the graph kernel.” Beamer et. al. Indistributed graph processing, communication (shuffles) dominate execution time.
46. Building a model for s Model Builder Input features Model Use machine learning to build a model. Learn the relation between s and error / latency
47. Building a model for s Model Builder Learn H: (s, a, g) => e/l Input features Model Use machine learning to build a model. Learn the relation between s and error / latency
48. Building a model for s Model Builder Learn H: (s, a, g) => e/l Input features Model Random Forests Use machine learning to build a model. Learn the relation between s and error / latency
49. Building a model for s Model Builder
50. Building a model for s Model Builder Input Graph
51. Building a model for s Model Builder Models Input Graph
52. Building a model for s Model Builder Models Input Graph Benchmark Graphs/Queries (e.g., Graph500)
53. Building a model for s Model Builder 0 1 4 2 3 Models Input Graph Benchmark Graphs/Queries (e.g., Graph500)
54. Building a model for s Model Builder Model Mapper 0 1 4 2 3 Models Input Graph Benchmark Graphs/Queries (e.g., Graph500)
55. Building a model for s Model Builder Model Mapper 0 1 4 2 3 Models Input Graph Benchmark Graphs/Queries (e.g., Graph500)
56. Building a model for s Model Builder Model Mapper 0 1 4 2 3 Models Input Graph Benchmark Graphs/Queries (e.g., Graph500)
57. Building a model for s Model Builder Model Mapper 0 1 4 2 3 Model Models Input Graph Benchmark Graphs/Queries (e.g., Graph500)
58. Preliminary Feasibility Evaluation § Implemented sparsifier on Apache Spark § Not limited to it § Evaluated on a few real-world graphs § Largest: UK 3.73B edges § Goal: check if our assumptions hold
59. Preliminary Feasibility Evaluation 0 0.5 1 1.5 2 2.5 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 Speedup Sparsifcation Parameter AstroPh Facebook 0 0.5 1 1.5 2 2.5 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 Speedup Sparsifcation Parameter Epinions Wikivote
60. Preliminary Feasibility Evaluation 0 0.5 1 1.5 2 2.5 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 Speedup Sparsifcation Parameter AstroPh Facebook 0 0.5 1 1.5 2 2.5 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 Speedup Sparsifcation Parameter Epinions Wikivote Performance trends similar for graphs that are similar
61. Preliminary Feasibility Evaluation � ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ���� ��� ���� ��� ���� ��� ���� ��� ���� ��� ���� ��� ���� ��� ���� ��� ������ ���������� ��� ������������� ��������� ������� �������� �������� ��������
62. Preliminary Feasibility Evaluation 1 2 3 4 5 6 7 8 9 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 20 30 40 50 60 70 Speedup Error (%) Sparsifcation Parameter Speedup Error
63. Preliminary Feasibility Evaluation 1 2 3 4 5 6 7 8 9 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 20 30 40 50 60 70 Speedup Error (%) Sparsifcation Parameter Speedup Error Bigger benefits achievable in large graphs
64. Ongoing/Future Work § Deep Learning § Build better models. § Better sparsifiers § Can we cherry pick sparsifiers? § Programming Language techniques § Can we synthesize approximate versions of an exact graph-parallel program?
65. Conclusion § Approximate graph analytics challenging § Unlike approximate query processing, no direct relation between graph size and latency/error. § Our proposal GAP: § Uses sparsification theory to reduce input to graph algorithms, and ML to learn the relation between input latency/error. § Initial results are encouraging. http://www.cs.berkeley.edu/~api api@cs.berkeley.edu