Better Fault Tolerance via ApplicationEnhanced Networks

1
Better Fault Tolerance via Application-Enhanced
Networks

• John Hartman
• Will Evans
• University of Arizona

2
End-to-end fault tolerance

• Fault tolerance as a QoS issue
• a broken QoS guarantee is a fault
• Denial-of-service is the tricky part of QoS
• accounting, detection, revocation
• while minimizing resource usage
• End-to-end argument
• DoS can happen at any interface
• application must be involved

3
Active networks

• Programmable network routers
• Use active network technologies to
• perform early prevention of DoS attacks
• tolerate faults (including DoS attacks)
• Application-enhanced networks

4
Project goals

• Develop local resource management for active
  routers
• fine-grain accounting and management
• necessary for tolerating faults and attacks
• Construct application-enhanced networks
• application-specific fault/attack prevention and
  response
• proofs-of-concept distributed terrain
  navigation, network-resident storage

5
Network-resident storage (NRS)

• Storage system functionality in network
• improve storage system performance and
  functionality
• centralize functionality in network
• storage-specific resource management and
  accounting
• e.g. quotas, permissions, bandwidth
• improve resistance to faults/attacks
• e.g. early reject of invalid requests

6
Active token service

• Initial NRS effort
• Token-based, network-resident synchronization
• Storage service-specific policies
• precedence, recovery
• early prevention of DoS attacks

7
Distributed terrain navigation
C
A
B
8
Fault prevention and tolerance

• Network links nodes fail
• Denial of service
• Prevention
• Detect rogue application clients
• Reject early
• Tolerance
• Application modifies network
• Network CPU storage compensate
  bandwidth/latency loss

9
Main tactics

• Sensitivity to data and network distance
• Joint compression
• Flexible multicast
• Flow aggregation
• Prediction
• Data migration

10
Use of active networks

• Application driven re-routing
• Compression at internal nodes
• Data migration