Transcript Slides in Powerpoint

An Efficient Topology-Adaptive
Membership Protocol for LargeScale Cluster-Based Services
§
Jingyu Zhou * , Lingkun Chu*, Tao Yang*
* Ask Jeeves
§University of California at Santa Barbara
§
Outline
Background & motivation
Membership protocol design
Implementation
Evaluation
Related work
Conclusion
Background
 Large-scale 24x7 Internet services
Thousands of machines connected by many level-2
and level-3 switches (e.g. 10,000 at Ask Jeeves)
Multi-tiered architecture with data partitioning and
replication
 Some of machines are unavailable frequently due
to failures, operational errors, and scheduled
service update.
Network Topology in Service Clusters
 Multiple hosting
centers across
Internet
 In a hosting center
 Thousands of nodes
 Many level-2 and
level-3 switches
 Complex switch
topology
Data
Center
Asia
Asian user
NY user
CA user
VP
N
N
3DNS
-WAN
Load
Balancer
P
V
Data
Center
California
VPN
Internet
Level-3 Switch
Level-2 Switch
Data
Center
New York
Level-2 Switch
Level-3 Switch
Level-2 Switch
...
Level-2 Switch
Motivation
Membership protocol
Yellow page directory – discovery of services
and their attributes
Server aliveness – quick fault detection
Challenges
Efficiency
Scalability
Fast detection
Fast Failure Detection is crucial
 Online auction service even with
Replica
1
replication
Failure of one replica 7s - 12s
Service unavailable 10s - 13s
Auction
Service
Replica
2
Replica
3
Communication Cost for Fast Detection
 Communication
requirement
Propagate to all nodes
Fast detection needs
higher packet rate
High bandwidth
 Higher hardware cost
 More chances of
failures.
Design Requirements of Membership
Protocol for Large-scale Clusters
Efficient: bandwidth, # of packets
Topology-adaptive: localize traffic within
switches
Scalable: scale to tens of thousands of nodes
Fast failure detection and information
propagation.
Approaches
Centralized
Easy to implement
Single point of failure, not scalable, extra delay
Distributed
All-to-all broadcast [Shen’01]: doesn’t scale well
Gossip [Renesse’98]: probabilistic guarantee
Ring: slow to handle multi-failures
Don’t consider network topology
TAMP: Topology-Adaptive Membership
Protocol
 Topology-awareness
Form a hierarchical tree according to network topology
 Topology-adaptiveness
Network changes: add/remove/move switches
Service changes: add/remove/move nodes
Exploit TTL field in IP packet
Hierarchical Tree Formation Algorithm
1. Form small multicast groups with low
TTL values;
2. Each multicast group performs elections;
3. Group leaders form higher level groups
with larger TTL values;
4. Stop when max. TTL value is reached;
otherwise, goto Step 2.
An Example
 3 Level-3 switches with 9
nodes
Group 3a
239.255.0.23
TTL=4
B
Group 2a
Group 2b
239.255.0.22239.255.0.22
TTL=3
TTL=3
B
A
A
B
C
C
Group 1a
239.255.0.21
TTL=2
Group 1b
239.255.0.21
TTL=2
Group 1c
239.255.0.21
TTL=2
A
B
C
Group 0a
239.255.0.20
TTL=1
Group 0b
239.255.0.20
TTL=1
Group 0c
239.255.0.20
TTL=1
A
B
C
Node Joining Procedure
 Purpose
 Find/elect a leader
 Exchange membership information
 Process
1. Join a channel and listen;
2. If a leader exists, stop and bootstrap with the
leader;
3. Otherwise, elects a leader (bully algorithm);
4. If is leader, increase channel ID & TTL, goto 1.
Properties of TAMP
 Upward propagation guarantee
A node is always aware of its leader
Messages can always be propagated to nodes in
the higher levels
 Downward propagation guarantee
A node at level i must leaders of level i-1, i-2, …, 0
Messages can always be propagated to lower level
nodes
 Eventual convergence
View of every node converges
Update protocol when cluster structure
changes
 Heartbeat for failure
detection
 Leader receive an update
- multicast up & down
3
Level 2
E
Level 1
2
2
B
E
2
3
Level 0
A
ABC
DEF
GHI
H
B
1
C
AB
DEF
GHI
D
DEF
ABC
GHI
3
E
F
4
DEF
AB
GHI
G
GHI
ABC
DEF
H
I
4
GHI
AB
DEF
Fault Tolerance Techniques
Leader failure: backup leader or election
Network partition failure
Timeout all nodes managed by a failed leader
Hierarchical timeout: longer timeout for higher
levels
Packet loss
Leaders exchanges deltas since last update
Piggyback last three changes
Scalability Analysis
 Protocols: all-to-all, gossip, and TAMP
 Basic performance factors
Failure detection time (Tfail_detect)
View convergence time (Tconverge)
Communication cost in terms of bandwidth (B)
Scalability Analysis (Cont.)
 Two metrics
BDP = B * Tfail_detect , lower failure detection time with
low bandwidth is desired
BCP = B * Tconverge , lower convergence time with low
bandwidth is desired
BDP
All-to-all
Gossip
TAMP
BCP
O(n2)
O(n2)
O(n2logn)
O(n2logn)
O(n)
O(n)+O(B*logkn)
n: total # of nodes
k: each group size, a constant
Implementation
 Inside Neptune middleware [Shen’01] –
programming and runtime support for
building cluster-based Internet services
 Can be easily coupled into others clustering
frameworks
Hierarchical Membership Service
Client
Code
SHM
External
Receiv
er
SHM
Local
Status
Tracker
Service
Status
Data
Structure
Inform
er
Conten
der
Annou
cer
Multicast Channels
Service
Code
/proc File
System
Evaluation: Objectives & Settings
 Metrics
 Bandwidth
 failure detection time
 View convergence time
 Hardware settings
 100 dual PIII 1.4GHz nodes
 2 switches connected by a Gigabit switch
 Protocol related settings
 Frequency: 1 packet/s
 A node is deemed dead after 5 consecutive loss
 Gossip mistake probability 0.1%
 # of nodes: 20 – 100 in step of 20
Bandwidth Consumption
 All-to-All & Gossip: quadratic increase
 TAMP: close to linear
Failure Detection Time
 Gossip: log(N) increase
 All-to-All & TAMP: constant
View Convergence Time
 Gossip: log(N) increase
 All-to-All & TAMP: constant
Related Work
 Membership & failure detection
[Chandra’96], [Fetzer’99], [Fetzer’01], [Neiger’96], and
[Stok’94]
 Gossip-style protocols
SCAMP, [Kempe’01], and [Renesse’98]
 High-availability system (e.g., HA-Linux, Linux
Heartbeat)
 Cluster-based network services
TACC, Porcupine, Neptune, Ninja
 Resource monitoring: Ganglia, NWS, MDS2
Contributions & Conclusions
TAMP is a highly efficient and scalable
protocol for giant clusters
Exploiting TTL count in IP packet for
topology-adaptive design.
Verified through property analysis and
experimentation.
Deployed at Ask Jeeves clusters with
thousands of machines.
Questions?