Jean Meloche

download report

Transcript Jean Meloche

Network Research
at
Avaya Labs Research
The ExpertNet and Chatter Teams
Akshay Adhikari
Mark Bearden
Lorraine Denby
Bengi Karacali
Jean Meloche
Balaji Rao
David Stott
Among many others
Challenges
• Network topology discovery
• Heisenberg
– Probes load the network at
the worst time
– Path or end to end
performance but not both
• Visualizations for paths
– Topology and paths
– Route changes
• Synchronization and One way
delay
• Load balancing and packet loss
• Missing values and clean up
• Complexity of networks
• Volume of data (GB per day)
Discovery via the fusion of
multiple data sources
•
•
•
•
•
•
SNMP
IP RR
Traceroute
I/O octet count correlation
Delay distance
End to end shadow of network events
Synchronization
Link level network delay via
TTL flood
• Multiple packets with IP RR for a range of
TTL are sent between endpoints
• ICMP messages are received from routers
along the way
• Results are used to estimate link level
delays
n 1
n
n 1
n
 (L   )   (L   )  L     
i 1
i
i
i 1
i
i
n
i 1
i
i 1
i
Network Sampling Design
• Where to place the probes
– Minimize the number of probes needed
– Minimize the degeneracy
– Subject to many constraints
• Where to measure QoS attributes and covariates
– Minimize the footprint
– Minimize the time to blame
• Dynamic sampling programs
Chatter
•
•
•
•
Distributed network monitor
Lives on cheap 100 MHz box
Online root cause analysis
Seeks failure patterns among arbitrary
covariates that are not specified a priori
• Uses adaptive control charts on important
QoS attributes to identify interesting
conditions
Compression
• Compression is needed to prevent the
transmission and storage of uninteresting
data
• The data are highly dimensional
• The collection is distributed
• How should the compression be distributed
Localization
• Data collection is distributed
• What methods are local (analysis requiring
only the local data) or near local (analysis
requiring limited exchange of information
between neighbors)?
• Are there sampling design that affect
localization?