Packet Scheduling in Ad hoc Networks

1
Packet Scheduling in Ad hoc Networks

• Mary Baker
• Byung-Gon Chun
• mgbaker_at_cs.stanford.edu
• http//mosquitonet.stanford.edu
• Stanford University

2
Packet scheduling - background

• Scheduling algorithm determines which packet
  queued at a node will be serviced next
• Used on switches, routers, etc.
• Scheduler affects end-to-end delay of packets

3
Why look at this in an ad hoc network?

• Scheduler performance may be different in ad hoc
  networks (compared to static wired networks) due
  to
• Frequent transmission of control packets
• Multi-hop forwarding of packets
• Nodes operate as routers, sources, and sinks of
  data
• Current default schedulers
• Give priority of control packets over data
  packets
• Do not differentiate between different data
  packets

4
Project goals

• Examine queuing behavior in ad hoc networks
• Experiment with different packet schedulers
• Should we always give priority to control
  packets?
• Should we differentiate between data packets?
• Look at two very different protocols
• DSR On demand non-geographic routing protocol
• GPSR proactive geographic routing protocol
• Important because ad hoc networks need all the
  performance help they can get

5
Routing protocol characteristics DSR

• DSR (dynamic source routing)
• On-demand cache routes, if not in cache when
  needed, ask for one
• Source routing route included in data packets as
  route header
• Route discovery
• A sends to B
• Route to B already in As cache?
• If not, broadcast RREQ packet with address of A
• Each node along the way appends its address to
  header
• Destination node appends its address and sends
  RREP
• Route maintenance
• If fatal error or no ack from downstream,
  generates RERR
• Remove routes from cache that use failed link
• Nodes back to source node do this

6
Routing protocol characteristics GPSR

• GPSR (greedy perimeter stateless routing)
• Nodes have GPS or such and broadcast positions
• Forward to neighbor closer to destination
• If no neighbor closer, switch to perimeter
  forwarding mode around hole
• Until a neighbor is closer to destination
• Switch back to geographic forwarding

7
Experimental setup

• Usual setup with ns-2 simulator
• Constant bit rate traffic
• Hidden waypoint model of mobility
• 50 nodes in a grid of 1500m x 300m
• 802.11 MAC layer and data link layer, 2 Mb/sec
  capacity
• Vary traffic load
• Change number of sources or packet sending rate
• Vary degree of mobility
• Vary pause times
• 0 and 900 reported
• Scheduler buffer size of 64 packets
• Send buffer size of 64 in DSR
• Packets dropped after 30 seconds in send buffer
• GPSR beacon set at 1 second

8
Results queuing behavior

• High traffic load
• A lot of queuing in both protocols
• Varying mobility
• Low queuing strikes particular regions
• High queue lengths mostly even across the
  network
• Composition of packets in queues
• Low mobility most packets in queue are data
• High mobility composition depends on protocol
• DSR has more routing packets than data packets
  often
• Due to RREP floods from many nodes with cached
  routes
• GPSR has mostly data packets

9
Types of schedulers studied

• Priority versus no-priority for control packets
• Different buffer management for each as well
• Round robin
• One packet per flow at a time (equal service to
  flows)
• Greedy
• Nodes sends its own data before others data
• Fewest hops remaining
• Packets sorted into classes by hops remaining
• Actually weighted round robin amongst classes to
  avoid starvation
• Weighted-distance scheduling as well for GPSR

10
Results scheduler

• Give priority to control packets?
• DSR with high mobility 40 decrease in average
  delay
• GPSR with high mobility 71 increase in average
  delay!
• Should we set priorities amongst data packets?
  Yes.
• New shortest distance schedulers work best
• DSR fewest remaining hops
• Decreases average delay by up to 32
• GPSR shortest remaining distance
• Decreases average delay up to 32
• Packets close to destination unlikely to suffer
  as much delay or interference

11
More details

• With higher sending rates we see a bigger
  difference
• With moderate to low mobility, more impact
• With GPSR, we see a difference even with high
  mobility
• Mostly data packets in queues

12
Conclusions, impact, future work

• Conclusion
• Ad hoc networks are different enough to warrant
  reinvestigating many of our assumptions
• Impact
• Forwarded to MC2R as a best poster paper from
  MobiHoc 2002
• Student Byung-Gon Chun accepted to UCB Ph.D.
  program
• Future work
• Would like to test this on AODV as well
• With help from Prof. Belding-Royer
• Try different traffic and mobility models
• Different simulation environment
• With help from Prof. Hou
• How many of these techniques hold for overlay
  networks?