Latency

1
Latency Scaling Issues in Mobile-IP

• Sreedhar Mukkamalla
• Bhaskaran Raman

2
Motivation

• ICEBERG project
• packet switched cellular telephony network
• flexible service architecture
• Universal INBOX
• Location based services etc.
• In search of the right mobility architecture

3
Project Goals

• Mobile IP IETF Standard for mobility
• Is it suitable for ICEBERG ?
• Wide area registration latency
• Is it a problem ?
• Large number of cell-phone users
• How many mobile hosts can an agent handle ?

4
Basic Mobile IP

• Triangle routing problem
• Wide-area registration latency
• Registration mechanism
• MH registers with HA through FA
• If registration times out, retry with exponential
  back-off
• Minimum RTO 1 second

HA
IP tunnel
CH
FA
Mobile Host
5
Wide area registration latency vs Time

• Ping measurements to Rutgers - used in simulation
• No congestion 70 ms latency During congestion
  1600 ms latency
• Congestion during day time 7am-5pm PST RTT
  540 ms
• With no back-off during losses, 1200 ms average
  latency

6
But...

• Machine at CMU
• average latency around 190 ms
• latency gt 200 ms only 3 of the time
• Machines at MIT UNC
• average lt 100 ms
• latency gt 200 ms less than 0.2 of the time
• Machine at UMD
• average 1200 ms
• latency gt 1 sec for more than 1/6th of the time

7
Scaling Measurements

• Experimental Setup
• FreeBSD implementation of Mobile-IP
• Pentium (300MHz) machines on 100 Mbps Ethernet
• Mobility emulated by modifying code at MH
• Measurements to determine
• Hand-off rate that an agent can handle
• Data traffic that an agent can handle

8
Results

• FA can handle about 2500 registrations per second
• User-level handling of registration messages
• HA/FA can handle data traffic (tunneling/de-tunnel
  ing) at 20,000 packets per second
• 160 byte packets
• Equivalent to about 1000 two-way audio streams
• Ethernet gets saturated
• Conclusion An agent can handle at least a few
  hundred mobile hosts

9
Conclusions

• Processing is not the bottleneck
• at least for up to a few 100 users per agent
• Wide-area registration latency
• Very less for some sites, unacceptably high for
  others
• Should we optimize the architecture at the IP
  layer for performance ?
• End-to-End argument gt IP mobility support
  required
• But, optimizations should probably be made at the
  link layer