Distributed Dynamic Replica Placement and Request Redirection in

Distributed Dynamic Replica Placement and Request Redirection in Content Delivery Networks Advisor : Ho-Ting Wu Student : Yu-Chiang Lin Date: 2011/5/30

OUTLINE p CDN introduce p Distributed Dynamic Replica Placement p Cloning of a replica p Replica removal p Redirection p Future work p Reference Page 2

OUTLINE p CDN introduce p Distributed Dynamic Replica Placement p Cloning of a replica p Replica removal p Redirection p Future work p Reference Page 3

CDN Introduce Fig: Server Farm(source from EBU technical review) Page 4

CDN Introduce p Consider TCP transmittion，throughput may affect by latency or packet lost。 Page 5

CDN Introduce p The Content Delivery Networks (CDN) paradigm is based on the idea to transparently move third-party content closer to the users p content is replicated on CDN servers which are located close to the final users so that users perceive a better content access service. Page 6

CDN Introduce Fig: Content Delivery Network(source from EBU technical review) Page 7

CDN Introduce Fig: Content Delivery Network Page 8

CDN Introduce p Four Important technique p 1. Content route p 2. Content distribution p 3. Content store p 4. Content management Page 9

CDN Introduce p CDN issue p 1) Deciding the kind of content that should be hosted (replica placement) p 2) selecting the best replica for a given user p 3) designing mechanisms for transparent redirection of the users requests to the best replicas Page 10

OUTLINE p CDN introduce p Distributed Dynamic Replica Placement p Cloning of a replica p Replica removal p Redirection p Future work p Reference Page 11

Distributed Dynamic Replica p Distributed scheme to allocate and deallocate replicas, so that the user requests are satisfied while minimizing the CDN costs in a dynamic scenario p This scheme always accounts for the current replica placement, adding replicas or changing replica location only when needed. p Each site j ∈ VR autonomously decides on whether some of the replicas it stores should be cloned or removed. Page 12

Problem Formulation p VA: access nodes p VR: CDN servers sites p d(i, j): user (access node) i to a replica j p dmax: distance threshold p xi, c: the volume of user requests originated at node i ∈ VA for content c ∈ C p Umax, Umid, Ulow: load threshold Page 13

Cloning of a replica p Function add_replica(j, c) p 1: lj, c =i∈VA αij, c · xi, c p 2: while lj, c/rj, c− Umax > 0 p 3: best_served = 0 p 4: best_distance = ∞ p 5: best_vr = undefined p 6: for all j’ ∈ ρ(j) s. t. rj, c < Vmax. R do Page 14

Cloning of a replica p 7: l’j, c =i∈α(j’) αij, c · xi, c p 8: total_distance =i∈α(j’)αij, c · xi, c · di, j’ p 9: if (l’j’, c < best_served) ∨ p 10: (l’j’, c = best_served ∧ p 11: total_distance < best_distance) then p 12: best_distance = total_distance p 13: best_served = l’j’, c Page 15

Cloning of a replica p 14: best_vr = j’ p 15: end if p 16: end for p 17: if best_vr = undefined then p 18: exit p 19: end if p 20: ask best_vr to add a replica Page 16

Cloning of a replica p 21: compute l’’bestvr, c = min(i∈α(bestvr) αij, c · xi, c, 1) p 22: lj, c = lj, c − l’’bestvr, c p 23: remove from the set of requests those that can be offloaded p 24: end while Page 17

Replica removal p A replica can be removed if (and only if) it serves no requests p A replica is removed only when it has not been serving requests for a time long enough to bring the exponential average down to zero. Page 18

Redirection p Based on this feedback users requests are directed away from a replica in response to threshold events (if the replica load exceeds Umax or falls below Ulow) p As an example, an underloaded replica informs the redirection system which then tries to offload requests to some other replicas (if possible) p A perfect load balancing may be impossible due to the distance constraint Page 19

Redirection Fig. A model for the redirection scheme Page 20

OUTLINE p CDN introduce p Distributed Dynamic Replica Placement p Cloning of a replica p Replica removal p Redirection p Future work p Reference Page 21

Future Work p comparison of different settings of the Umid p observe dmax and add remove replica relationship Page 22

Reference p [1] N. Bartolini, F. Lo Presti, and C. Petrioli, “Optimal dynamic replica placement in Content Delivery Networks, ” in Proceedings of the 11 th IEEE International Conference on Networks, ICON 2003, Sydney, Australia, September 28– October 1 2003, pp. 125– 130. p [2] F. Lo Presti, C. Petrioli, and C. Vicari, “Dynamic replica placement in content delivery networks, ” in Proceedings of MASCOTS 05, September 2005. p [3] F. Lo Presti, C. Petrioli, and C. Vicari, “Distributed Dynamic Replica Placement and Request Redirection in Content Delivery Networks , ” in Proceedings of MASCOTS 07, 2008. Page 23