Create Presentation
Download Presentation

Download Presentation
## Performance Evaluation

- - - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - - -

**Performance Evaluation**Lecture 1: Complex Networks Giovanni Neglia INRIA – EPI Maestro 10 December 2012**Outline**• Properties of Complex Networks • Small diameter • High Clustering • Hubs and heavy tails • Physical causes • Consequences**Small Diameter (after Milgram’s experiment)**Six degrees - the science of a connected age, 2003 Six degrees of separation is the idea that everyone is on average approximately six steps away, by way of introduction, from any other person in the world, so that a chain of "a friend of a friend" statements can be made, on average, to connect any two people in six steps or fewer. J. Vaucher, Small World Networks, April 2005**Small Diameter, more formally**• A linear network has diameter N-1 and average distance Θ(N) • How to calculate it? • A square grid has diameter and average distance Θ(sqrt(N)) • Small World: diameter O((log(N))a), a>0 • Lessons from model: long distance random connections are enough**Erdös-Rényi graph**• A ER graph G(N,q) is a stochasticprocess • N nodes and edges are selectedwithprob. q • Purpose: abstract from the details of a given graph and reach conclusions depending on itsaveragefeatures**Erdös-Rényi graph**• A ER graph G(N,q) is a stochastic process • N nodes and edges are selected with prob. q • Degree distribution: P(d)=CdN-1 qd(1-q)N-1-d • Average degree: <d>=q (N-1) • For N->∞ and Nq constant: P(d)=e-<d><d>d/d! • <d2>=<d>(1+<d>) • Average distance: <l>≈logN/log<d> • Small world**Clustering**• "The friends of my friends are my friends" • Local clustering coefficient of node i • (# of closed triplets with i at the center) / (# of triplets with node i at the center) = (links among i’s neighbors of node i)/(potential links among i’s neighbors) • Global clustering coefficient • (total # of closed triplets)/(total # of triplets) • # of triplets = 3 # of triangles • Or 1/N Σi Ci Ci=2/(4*3/2)=1/3**Clustering**• In ER • C ≈ q ≈ <d>/N**Clustering**• In real networks Good matching for avg distance, Bad matching for clustering coefficient Ci=2/(4*3)**How to model real networks?**Regular Graphs have a high clustering coefficient but also a high diameter Random Graphs have a low diameter but a low clustering coefficient --> Combine both to model real networks: the Watts and Strogatz model Random Graph (k=4) Short path length • L=logkN Almost no clustering • C=k/n Regular Graph (k=4) Long paths • L = n/(2k) Highly clustered • C=3/4 Regular ring lattice R. Albert and A.-L. Barabasi: Statistical mechanics of complex networks, Rev. Mod. Phys., Vol. 74, No. 1, January 2002**Watts and Strogatz model**Random rewiring of regular graph With probability p rewire each link in a regular graph to a randomly selected node Resulting graph has properties, both of regular and random graphs --> High clustering and short path length R. Albert and A.-L. Barabasi: Statistical mechanics of complex networks, Rev. Mod. Phys., Vol. 74, No. 1, January 2002**Small World**• Usually to denote • smalldiameter + highclustering**Intermezzo: navigation**• In Small world nets there are short paths O((log(N))a) • But can we find them? • Milgram’s experiment suggests nodes can find them using only local information • Standard routing algorithms require O(N) information**Kleinberg’s result**• Model: Each node has • Short-range connections • 1 long-range connection, up to distance r with probability prop. to r-α • For α=0 it is similar to Watts-Strogatz model: there are short-paths**Kleinberg’s result**• If α=2 the greedy algorithm (forward the packet to the neighbor with position closest to the destination) achieves avg path length O((log(N))2)**Kleinberg’s result**• If α<>2 no local information algorithm can take advantage of small world properties • avg path length Ω(Nβ/2) • where β=(2-α)/3 for 0<=α<=2, β=(α-2)/(α-1), for α>2**Kleinberg’s result**• Conclusions • The larger α the less distant long-range contacts move the message, but the more nodes can take advantage of their “geographic structure” • α=2 achieved the best trade-off**Hubs**• 80/20 rule • few nodes with degree much higher than the average • a lot of nodes with degree smaller than the average • (imagine Bill Clinton enters this room, how rapresentative is the avg income) • ER with N=1000, <d>=5, P(d)≈e-<d><d>d/d! • #nodes with d=10: N*P(10)≈18 • #nodes with d=20: N*P(20)≈2.6 10-4 P(d) d**Hubs**ER Power law PDF CCDF Power law: P(d) ~ d-α**5: DataLink Layer**Power lawdegree distributions coauthorship**5: DataLink Layer**… and more Income Meme on Twitter Deaths in terroristic attacks**Power Law**• Where does it come from? • Albert-Barabasi’s growth model • Highly Optimized Model • And other models • See Michael Mitzenmacher, A Brief History of Generative Models for Power Law and Lognormal Distributions**Albert-Barabasi’s model**• Twoelements • Growth • m0 initial nodes, every time unit weadd a new nodewith m links to existingnodes • Preferentialattachment • The new node links to a nodewithdegreekiwithprobability**Albert-Barabasi’s model**• Node i arrives at time ti, itsdegreekeepsincreasing • With a continuum approximation: • Thendegree distribution at time tis:**Albert-Barabasi’s model**• At time tthere are m0+t nodes, if weconsiderthat the tnodes are addeduniformlyatrandom in [0,t], then**Albert-Barabasi’s model**• The PDF is • For t->∞**Albert-Barabasi’s model**• If , • Othervariants: • With fitness • Withrewiring (a prob. p to rewire an existingconnection) • Uniform attachingwith "aging": A vertex isdeactivatedwith a prob. proportional to (ki+a)-1**Back to Navigation: RandomWalks**• Whatcanwe do in networks without a geographical structure? • Randomwalks 1/4 1/4 1/2 1/4 1/4 1/2**Back to Navigation: RandomWalks**• How much time isneeded in order to reach a givennode?**RandomWalks: stationary distribution**• avg time to come back to node i starting from node i: • Avg time to reach node i • intuitively ≈Θ(M/ki) i**Another justification**• Randomwalk as randomedgesampling • Prob. to pick an edge (and a direction) leading to a node of degree k is • Prob. to arrive to a givennode of degree k: • Avg. time to arrive to thisnode 2M/k**Distributed navigation (speed up randomwalks)**• Everynodeknowsitsneighbors {a,b,c,d} d a i c b**Distributed navigation (speed up randomwalks)**• Everynodeknowsitsneighbors • If a randomwalklooking for i arrives in a the message isdirectlyforwarded to i {a,b,c,d} d a i c b**Distributed navigationreasoning 1**• We discover i when we sample one of the links of i’s neighbors • Avg # of these links: • Prob. to arrive at one of them: d i a c b**Distributed navigationreasoning 2**• Probthat a node of degree k isneighbor of nodei • Probthat the nextedgebrings to a nodethatisneighbor of node i:**Distributed navigation**• Avg. Hop# • Regular graph withdegree d: • ER with <k>: • Pareto distribution : If α->2…**Distributed navigation**• Application example: • File search in unstructured P2P networks throughRWs**Configuration model**• A family of random graphs withgivendegree distribution**Configuration model**• A family of random graphs withgivendegree distribution • Uniform randommatching of stubs**Configuration model**• A family of random graphs withgivendegree distribution • Uniform randommatching of stubs