Dynamic Replica Placement for Scalable Content Delivery

1
Dynamic Replica Placement for Scalable Content
Delivery
Yan Chen, Randy H. Katz, John D.
Kubiatowicz yanchen, randy, kubitron_at_CS.Berkeley
.EDU EECS Department UC Berkeley
2
Motivation Scenario
data source
data plane
Web content server
CDN server
network plane
3
Goal and Challenges
Provide content distribution to clients with good
Quality of Service (QoS) while retaining
efficient and balanced resource consumption of
the underlying infrastructure

• Dynamically choose the number and placement of
  replicas while satisfying clients QoS and
  servers capacity constraints
• Good performance for update dissemination
• Delay - Bandwidth consumption
• Without global network topology knowledge
• Scalability for millions of objects, clients and
  servers

4
Outline

• Goal and Challenges
• Previous Work
• Our Solutions Dissemination Tree
• Peer-to-peer Location Service
• Replica Placement and Tree Construction
• Evaluation
• Conclusion and Future Work

5
Previous Work

• Focused on static replica placement
• Clients distributions and access patterns known
  in advance
• Assume global IP network topology
• DNS-redirection based CDNs highly inefficient
• Centralized CDN name server cannot record replica
  locations
• No inter-domain IP multicast
• Application-level multicast (ALM) unscalable
• Root maintains states for all children or handle
  all join requests

6
Solution Dissemination Tree

• Dynamic replica placement use close-to-minimum
  number of replicas to satisfy QoS and capacity
  constraints with local network topology

• Adaptive cache coherence for efficient coherence
  notification

• Use Tapestry location service to improve the
  scalability locality

data source
data plane
root server
server
network plane
7
Peer-to-peer Routing and Location Services

• Properties Needed by d-tree
• Distributed, scalable location with guaranteed
  success
• Search with locality
• P2P Routing and Location Services
• CAN, Chord, Pastry, Tapestry, etc.
• Http//www.cs.berkeley.edu/ravenben/tapestry

8
Integrated Replica Placement and d-tree
Construction

• Dynamic Replica Placement Application-level
  Multicast
• Naïve approach - Smart approach
• Static Replica Placement IP Multicast
• Modeled as capacitated facility location problem
• Design a greedy algorithm with logN approximation
• Optimal case for comparison
• Soft-state Tree Maintenance

9
Dynamic Replica Placement naïve
data plane
s
surrogate
c
network plane
10
Dynamic Replica Placement naïve
data plane
parent candidate
s
surrogate
c
network plane
Tapestry overlay path
first placement choice
11
Dynamic Replica Placement smart
data plane
client child
s
parent
surrogate
sibling
c
server child
network plane
12
Dynamic Replica Placement smart

• Aggressive search
• Lazy placement

• Greedy load distribution

data plane
parent candidates
client child
s
parent
surrogate
sibling
c
server child
network plane
Tapestry overlay path
first placement choice
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
• Network Simulator
• NS-like packet-level priority-queue based event
  simulator
• 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
Performance of Dynamic Replica Placement

• Compare Four Approaches
• 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)
• Metrics
• Number of replicas deployed and load distribution
• Dissemination multicast performance
• Tree construction traffic

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

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 and Future Work

• Dissemination Tree dynamic Content Distribution
  Network on top of a peer-to-peer location service
• Dynamic replica placement satisfy QoS and
  capacity constraints and self-organize into an
  app-level multicast tree
• Use Tapestry to improve the scalability and
  locality
• Simulation Results Show
• Close to optimal number of replicas, good load
  distribution, low multicast delay and bandwidth
  penalty at the price of reasonable construction
  traffic
• Future Work
• Evaluate with more diverse topologies and
  workload
• Dynamic replica deletion/migration to adapt to
  the shift of users interests

19
Routing in Detail
Example Octal digits, 212 namespace, 5712 ? 7510
5712
0880
3210
4510
7510
20
Dynamic Replica Placement

• Client c sends join request to statistically
  closest server s with object o through nearest
  representative server rs
• Naïve placement only checks s before placing new
  replicas while smart algorithm in addition checks
  parent, free siblings and free server children of
  s
• Remaining capacity info piggybacked in soft-state
  messages
• If unsatisfied, s place new replica on one of the
  Tapestry path nodes from c to s (path piggybacked
  in the request)
• Naïve one always chooses the closest qualified
  node (to c), while smart one puts on the furthest
  qualified node

21
Static Replica Placement

• Without Capacity Constraints Minimal Set Cover
• Most cost-effective method Greedy algorithm
  Grossman Wool 1994
• With Capacity Constraints Variant of Capacitated
  Facility Location Problem
• C demand locations, S locations to build
  facilities, the capacity installed at each
  location is an integer multiple of u
• Facility building cost f_i, service cost c_ij
• Objective function minimize sum of f_i and c_ij
• Mapped to our problem f_i 1 and c_ij 0 if
  location i cover demand j and infinity otherwise

22
Static Replica Placement Solutions

• Best theoretical one use Primal-dual schema and
  Lagrangian relaxation Jain Varizani 1999
• Approximation ratio 4
• Variant of greedy algorithm
• Approximation ratio lnS
• Choose s with the largest value of
  min(cardinality C_s, remaining capacity RC_s)
• If RC_s lt C_s, choose those least-covered
  clients to cover first
• With Global IP topology vs. with Tapestry overlay
  path topology only

23
Soft State Tree Maintenance

• Bi-directional Messaging
• Heartbeat message downstream refresh message
  upstream
• Scalability
• Each member only maintains states for direct
  childrenparent
• Join request can be handled by any member
• Fault-tolerance through rejoining

24
Performance Under Various Conditions

• With 100, 1000 or 4500 clients
• With or without load constraints
• Od_smart performs consistently better than
  od_naïve and close to IP_s as before
• Load constraint can avoid hot spot

25
Performance Under Various Conditions II

• With 2500 random d-tree servers
• Merit of scalability maximal number of server
  children is very small compared with total number
  of servers
• Due to the randomized and search with locality
  properties of Tapestry

26
(No Transcript)