SuDuelKu
Fault Tolerant Distributed Systems
6
8
1
4
3
5
5
6
2
1
8
4
7
6
7
6
3
9
1
4
5
2
7
4
3/22/2006
2
6
5
8
9
7
SuDuelKu - Team1
Team 1
Saul Jaspan
Yudi Nagata
Lucia de Lascurain
Chris Nelson
Luis Rios
1
Agenda
 Application
Configuration
 Fault Tolerant Architecture
 Questions and Assumptions
 Lessons learned
 What’s next
3/22/2006
SuDuelKu - Team1
2
Application Configuration
 EJB
 Linux
 Java
 Using:
– JBoss
– MySQL
– Lomboz
– Eclipse
3/22/2006
SuDuelKu - Team1
3
Fault Tolerant Architecture - 1
Fault
Injector
Name Server
Client 1
Server 1
…
Database
Client N
Server 2
(backup)
3/22/2006
SuDuelKu - Team1
Recovery
Manager
4
Fault Tolerant Architecture - 2
Name
Server
SuDuelKuClient
primary
Recovery Manager
SuDuelKu
Server
backup
void startNewServer(…)
primary
backup
EJBClient
FaultTolerant SuDuelKu Server
FaultInjecting EJBClient
void checkPoint(…)
bool isAlive(…)
void setPacketDropRate(…)
void setCheckpointFrequency(…)
void setAsPrimary(…)
void addBackup(…)
void addCommand(…)
void setPacketDropRate(…)
void setCommandRate(…)
primary
FaultInjector
void
void
void
void
startNewClient(…)
setPacketDropRate(…)
setLoad(…)
setErrorTypes(…)
3/22/2006
backup
CommandQueue
SuDuelKu - Team1
void addCommand(…)
void clearQueue(…)
void getCommands(…)
5
Questions and Assumptions
 Questions:
– Are we really 100% stateless?
– Do we need checkpointing if we are stateless?
– Should we have a separate JNDI?
 Assumptions:
– Our naming server will not be killed.
– Our message IDs will be committed to the
database with the transactions atomically.
3/22/2006
SuDuelKu - Team1
6
Lessons Learned
 EJB
learning curve steep but eases
development once summited
 Calling back from the server to the client
is unnatural with middleware
 Fault Recovery is MUCH easier when the
middle tier is stateless
3/22/2006
SuDuelKu - Team1
7
What’s next
 Develop
fault tolerant system.
 Develop fault injector.
 Obtain system metrics.
3/22/2006
SuDuelKu - Team1
8
Questions
3/22/2006
SuDuelKu - Team1
9