Cooperative Computing for Autonomous Data Centers

Size: px

Start display at page:

Download "Cooperative Computing for Autonomous Data Centers"

Rodney Baker
6 years ago
Views:

1 Cooperative Computing for Autonomous Data Centers Jared Saia (U. New Mexico) Joint with: Jon Berry, Cindy Phillips (Sandia Labs), Aaron Kearns (U. New Mexico)

2 Outline 1) Bounding clustering coefficients in social networks 2) An application: Cooperative Computing

3 Power Law Bounded (PLB) Assumption [Brach, et al. 16] The PLB assumption holds for a network if, for some α > 1: The number of vertices with degree k is O(n k 1 - α )

4 Power Law Bounded Neighborhood (PLBN) Assumption [Brach, et al. 15] The PLBN Assumption holds for a network if for some α, 2 < α < 3: For every vertex v, of degree d, the number of neighbors of v of degree d is: O(d 3 - α )

5 Edge Weights Consider an edge (u,v) Let dᵤ, dᵥ be the degrees of u, v Let t be the number of triangles containing (u,v) Edge Weight: 2t / (dᵤ + dᵥ - 2)

6 Bounded Weighted Degree (BWD) Assumption For any human node, the sum of the weights of edges on that node is c, where c is a universal constant Inspired by Dunbar-like argument

7 Weighted Degrees Probability Density Extremely Unlikely any vertex has weighted degree above 40 Weighted Degree

8 Edge Weights Revisited Let dᵤ, dᵥ be the degrees of u, v Let t be the number of triangles containing (u,v) Edge Weight 1: 2t / (dᵤ + dᵥ - 2) Edge Weight 2: t / (dᵤ + dᵥ - t) Both of these greater than: t / (dᵤ + dᵥ)

9 O(d 4-α ) triangle bound Assume: PLBN: Number of neighbors of v with degree d is O(d 3 - α ) BWD: Sum of edge weights touching v is O(1) Then: Number of triangles v is in is O(d 4 - α )

10 Proof Sketch (1) Neighbors with degree d Each is in O(d) triangles with v O(d 3 - α ) of them (by PLBN) Contribute O(d 4 - α ) triangles

11 Proof Sketch (2) Neighbors with degree < d Let x = # of triangles these nodes contribute Claim: Sum of edge weights is minimized when x is evenly distributed Then, sum of edge weights is: Ω(d((x/d)/d)) = Ω(x/d) By the BWD Assumption, this sum is O(1). Thus x = O(d)

12 Clustering Coefficient Consider a node v, with degree d Clustering coefficient of v: # of triangles containing v / (d choose 2) Prediction: clustering coefficient is O(d 2-α )

13 Cleaning Networks

14 Removing Non-human Nodes Can buy 500 followers for $5 on Twitter Non-human nodes can be densely connected Our goal: remove non-human nodes

15 Removing Non-human Nodes Our approach (sketch): Remove nodes based on significant deviation from average weighted degree Cross-validate with human labelling and labelling based on non-topological features Cleaning removes few human nodes, among those we checked by hand J. Berry, A. Kearns, C. Phillips, J. Saia. Finding Non-human Nodes in Electronic Social Networks Using Only Topological Data Under Submission, 2016.

16 Does our Bound on Clustering Coefficients Hold on Cleaned Networks?

17 Youtube and Twitter

18 1e+06 Power Law Bounds on Degree Distributions Frequency Youtube (cleaned 6-σ) Twitter (cleaned 6-σ) α = 2.2 α = Degree 1 Clustering Coefficients Clustering Coefficient Youtube Twitter α' = Degree

19 LiveJournal and Friendster

20 1e+06 Power Law Bounds on Degree Distributions Frequency LiveJournal Frienster α = 2.5 α = Degree 1 Clustering Coefficients Clustering Coefficient LiveJournal (cleaned 6-σ) Frienster α' = 0.5 α' = Degree

21 The Takeaway Preliminary experiments suggest clustering coefficient is O(d 2-α ) Hidden constant seems fairly small Not necessarily tight Only holds after cleaning

22 Outline 1) Bounding clustering coefficients in social networks 2) An application: Cooperative Computing

23 Cooperative Computing

24 Cooperative Computing Alice and Bob (or more) have graphs G A and G B Share the name space for nodes Cooperate to solve problems on G A G B Goal: limit total communication s

25 Previous Work: Algorithms for s-t connectivity Assumes social network structure: giant component O(log² n) communication cost Can ensure players get no information beyond the output in honest-but-curious model J. Berry, J, M. Collins, A. Kearns, C. Phillips, J. Saia, and R. Smith. "Cooperative Computing for Autonomous Data Centers." IEEE Parallel and Distributed Processing Symposium (IPDPS), 2015.

26 CCR Finding a clique in G A G B Start with a social network O(log n) nodes chosen randomly to form a clique Adversary assigns edges to Alice and Bob Goal: Find the clique with little communication Blue edges to Alice Red edges to Bob

27 CCR Assumptions 1) PLB: for some α > 2, the number of vertices with degree k is O(n k 1 - α ) 2) PLBN + BWD: for some α > 2, any degree d nodes is in O(d 4 - α ) triangles Idea: Use these assumptions to bound the number of triangles outside the clique Blue edges to Alice Red edges to Bob

28 CCR Step 1: Ramsey Theory Plant a clique of size x, and split the edges arbitrarily between Alice and Bob Ramsey theory: either Alice or Bob has θ(x³) of the triangles Blue edges to Alice Red edges to Bob

29 Step 2: Find Maximum Triangle Dense Subgraph Goal: Find a subgraph that maximizes triangle density: # triangles in subgraph / # nodes in subgraph Triangle density is: 7/5 Can solve in polynomial time using Linear Programming Greedy 3-approximation (from Charikar s 2- approximation result for maximizing edge density)

30 Step 3: Communicate WLOG, assume Alice is the one who has θ(log³ n) triangles from the planted clique Lemma: Alice s maximum triangle dense subgraph contains only nodes in the planted clique, S Lemma: Whp, any nodes not in S have O(1) edges into S

31 Algorithm 1.Alice finds max triangle-density subgraph, H, and nodes (W A ) adjacent to at least half of H. Sends H and W A to Bob. 2.Bob finds nodes (W B ) adjacent to at least half of H, and sends all induced edges (between H, W A and W B ) 3.Alice finds clique (polynomial-time since O(log n)) H W A Alice s MTDS H W A W B W B

32 Main Result Can cooperatively find a O(log n) size clique planted in a social network Requires only polylogarithmic communication Key Assumptions: Power law network Bounded weighted degree

33 Questions?

34 Extra Slides

35 Planted Clique Problem Find a clique that has been artificially added to a graph O(log n) nodes chosen randomly and build a clique Can we find a clique that s a little larger than native clique size? For Erdos-Renyi, native is log n, can find (Deshpande and Montanari 2013, Alon, Krivelevich, Sudakov, 1998) A form of anomaly detection As a hardness assumption, hardness bounds for Finding a Nash equilibrium in a 2-party game (Hazan & Krauthgamer (2011)) Finding sparse principal components (Berthet and Rigollet (2013)) Testing whether distribution is k-wise independent (Alon et al 2007) NYC Theory Day April 15,

36 References Edge Weight 1: [J. Berry, B. Hendrickson, R. LaViolette, and C. Phillips, Tolerating the community detection resolution limit with edge weighting, Physical Review E, Vol. 83, No. 5, 2011.] Edge Weight 2: [Easley, David, and Jon Kleinberg. Networks, crowds, and markets: Reasoning about a highly connected world. Cambridge University Press, 2010.]

37 References PLB and PLBN Assumption: [Brach, P., Cygan, M., Łącki, J. and Sankowski, P., Algorithmic Complexity of Power Law Networks, Symposium on Discrete Algorithms (SODA), 2016.]

Distribution-Free Models of Social and Information Networks

Distribution-Free Models of Social and Information Networks Tim Roughgarden (Stanford CS) joint work with Jacob Fox (Stanford Math), Rishi Gupta (Stanford CS), C. Seshadhri (UC Santa Cruz), Fan Wei (Stanford