1. Dynamic Replica Placement for Scalable Content Delivery
Yan Chen, Randy H. Katz,
John D. Kubiatowicz
{yanchen, randy,
EECS Department
UC Berkeley

2. Motivation Scenario data source replica data plane cache always update
adaptive
coherence
always update
replica
data plane
client
CDN server
Web content
server
network plane

3. Goal and Challenges Dynamic choice of number and location of replicas
Provide content distribution to clients with good Quality of
Service (QoS) while retaining efficient and balanced resource
consumption of the underlying infrastructure
Dynamic choice of number and location of replicas
Clients’ QoS constraints
Servers’ capacity constraints
Efficient update dissemination
Delay
Bandwidth consumption
Scalability: millions of objects, clients and servers
No global network topology knowledge
We are going to compare several approaches, dynamic replica placement…
And how they are self-organized into d-tree and how they are maintained through soft-state mgt.

4. Previous Work (Replica Placement)
Focused on static replica placement
Clients’ distributions and access patterns known in advance
Assume global IP network topology
Data Location via DNS-redirection
Highly inefficient (this is a “hack”)
Centralized CDN name server cannot record replica locations

5. Previous Work (Info Dissemination)
No inter-domain IP multicast
Application-level multicast (ALM) unscalable
Root maintains states for all children (Narada, Overcast, ALMI, RMX)
Root handles all “join” requests (Bayeux)
Root split is common solution, but suffers consistency overhead

6. Solutions for Dissemination Tree
Peer-to-Peer Overlay Location Services with Good Scalability & Locality
Simultaneous Replica Placement and Tree Construction

7. Peer-to-peer Routing and Location Services
Properties Needed by Tree Building Algorithms
Distributed, scalable location with guaranteed success
Search with locality
P2P Routing and Location Services: Tapestry
CAN, Chord, Pastry insufficient locality or flexibility to place objects

8. Simultaneous Replica Placement and Tree Construction
Static Replica Placement + IP Multicast
Modeled as a global optimization problem
Design a greedy algorithm with logN approximation
Optimal case for comparison
Dynamic Replica Placement + Application-level Multicast
Search for qualified local replicas first
Place new replicas on Tapestry overlay path
Two approaches: naïve and smart
Soft-state Tree Maintenance
Each node only maintains states for its parent and direct children
We are going to compare several approaches, dynamic replica placement…
And how they are self-organized into d-tree and how they are maintained through soft-state mgt.
Talk about facility location problem:
Each CDN server is a facility and each client is a location, to put a replica incurs an uniform cost 1. Each CDN server can cover a certain subset of clients and has a capacity constraint on the # of clients to cover.
To minimize the total cost while covering all clients.

9. Dynamic Replica Placement: naïve
data plane
parent candidate
Tapestry overlay path s
proxy c
network plane
Tapestry mesh

10. Dynamic Replica Placement: naïve
data plane
parent candidate s
proxy
Piggyback the overlay dist and load of the servers on the Tapestry path, thus it will choose the one as close to c as possible c
first placement choice
network plane
Tapestry mesh
Tapestry overlay path

11. Dynamic Replica Placement: smart
parent candidates
Aggressive search
Greedy load distribution
data plane
client child
Tapestry overlay path
Again, client c submit request to s through Tapestry
S have a larger set, including …, we call it aggressive search…
Load-qualified candidates will send its load to c, c will ping them to get the distance and choose the one both satisfy the latency constraint and has the we choose the one with the lightest load as parent, we call it greedy load dist
Then it will place a cache on the client and tie it into the d-tree
If none of them satisfy, c will send a another lazy placement msg to s through Tapestry, piggyback the load of the servers on the Tapestry path.
S will use the overlay distance for estimation of IP distance, and choose the one as far to c as possible s
proxy
parent
sibling c
server child
network plane

12. Dynamic Replica Placement: smart
Aggressive search
Lazy placement
Greedy load distribution
data plane
parent candidates
client child
Tapestry overlay path s
proxy
parent
first placement choice
sibling c
server child
network plane

13. Evaluation Methodology
Network Topology
5000-node network with GT-ITM transit-stub model
500 d-tree server nodes, 4500 clients join in random order
Dissemination Tree Server Deployment
Random d-tree
Backbone d-tree (choose backbone routers and subnet gateways first)
Constraints
50 ms latency bound and 200 clients/server load bound

14. Four Approaches for Comparison
Overlay Dynamic Naïve Placement (dynamic_naïve)
Overlay Dynamic Smart Placement (dynamic_smart)
Static Placement on Overlay Network (overlay_static)
Static Placement on IP Network (IP_static)

15. Number of Replicas Deployed and Load Distribution
Overlay_smart uses much less replicas than overlay_naïve and very close to IP_static
Overlay_smart has better load distribution than od_naïve, overlay_static and very close to IP_static

16. Multicast Performance
85% of overlay_smart Relative Delay Penalty (RDP) less than 4
Bandwidth consumed by overlay_smart is very close to IP_static and much less than overlay_naive
With fixed replicas, how good multicast performance d-tree has

17. Tree Construction Traffic
Including “join” requests, “ping” messages, replica placement and parent/child registration
Overlay_smart consumes three to four times of traffic than overlay_naïve, and the traffic of overlay_naïve is quite close to IP_static
Far less frequent event than update dissemination

18. Conclusions Peer-to-peer networks can be used to construct CDNs
Dissemination Tree: dynamic Content Distribution Network with good QoS, efficiency and load balancing
P2P location service to improve scalability and locality
Simultaneous dynamic replica placement and tree construction
In particular
Use Tapestry to contact nearby region of tree to select parent
Lazy placement of new replicas on Tapestry overlay path
Close to optimal number of replicas, good load distribution, low multicast delay and bandwidth penalty at the price of reasonable construction traffic

19. Future Work Evaluate with more diverse topologies and real workload
Dynamic replica deletion/migration to adapt to the shift of users’ interests
Implementation for OceanStore, a global-scale persistent data storage system