karma

advertisement
Karthik Dantu
Bryan Kate
Jason Waterman
Peter Bailis
Matt Welsh
Presentors: Yuxuan Dai, Long Ma
Advantages
Extremely small
Large number
Can be applied to enclosed,
close-quarter areas
deployed to achieve a task
massively
Can
perform tasks that
challenging for larger platforms
Provide robustness to failure in
the field
weight buget
Restrictions
actuation
sensor and
others
Limited
resources
power buget
Complex
coordination
actuation
cumputating
and sensing
swarm size
Simplify
MAVs’
behavior
uniformly
Combine MAV
behaviors with
a swarm-level
goal
Eliminate low-level
MAV coordination
from users
programming
hive-drone model
model
drones cannot
communicate
with each other
(X, Y)
drones operate
without precise
knowledge of
location
Introduction
Hive-Drone Model
Karma Implementation
Evaluation
Our Thoughts
Coordination
complexity
Programmer cannot explicitly
coordinate MAV’s behavior
drones cannot communicate
with each other
1.Simiply MAV
programming
Introduce a
information
2.Better decision delay in the
making
system
Sortie
Disease!
Behavior
Application
Disease!
Spatial Decomposition
1. Make it easier to reason about MAV
allocation.
2. It is unlikely that the MAVs can access
to high-resolution location services in
the field
3. Using a Cartesian coordinate system
is not necessary
T,R,N
T,R,N
Data Model
The hive maintains a key-value
repository called the Datastore.
Updates to this data structure are
The Datastore can be queried both
asynchronous, occurring when
temporally and spatially
drones return from a sortie
Programming Model
Every behavior produces some type of
information under normal execution
1. Activation Predicate
2. Progress Function
Scheduling Problem
1. Use the shortest time to Complete
application
Drone 1 run
behaviors A in R1 (t1)
Drone 4 run behaviors
A in R3 (t4)
Advocate scheduling all
The hive-drone, programming model transform the
available drones greedily
problem of executing an application on a MAV swarm
into a problem of scheduling behaviors on drones.
Drone 2 run
Drone 3 run
System executes
behaviors
are
behaviors
B inthat
R6 behaviors
A in R7
concurrently activated in sequence
(t2)
(t3)
If behavior A and B are
activated concurrently
(greedily schedule all
available drones)
Interleaves allocations
for A and B
Schedule all drones for A
then for B
No distinction
Drone N run behaviors
Drone 1 run
A in Rx
behaviors A in R1
(tn)
(t1)
Drone 1 run
behaviors B in R1
(t2)
Drone n run
behaviors B in Rx
(tn)
Scheduling Problem
1. Use the shortest time to Complete
application
2. Achieve fairness between
behaviors
Minimize the difference in progress
between any two activated behaviors
Introduction
Hive-Drone Model
Karma Implementation
Evaluation
Our Thoughts
Karma
Controller
Scheduler
Karma Implementation
Dispatcher
Datastore
Process
Total
Workload
Remaining
Workload
Service
Level
Allocated
Drones
A
100
80
0.8
40
B
100
20
0.2
10
Process:
a. Estimate the total
workload
b. Allocate the available
drones fairly basing
the remaining amount
of work.
Process:
a. Help drove transmit
the data to hive
Datastore
b. Charge the droves
c. Notify the Controller
of the resource
availability
Two kinds of Dispatch policies:
a. Continuous dispatch policy
Provide a constant presence of drones in the field
Minimize the information latency
b. Greedy dispatch policy
Dispatch the drones opportunistically
Introduction
Hive-Drone Model
Karma Implementation
Evaluation
Our Thoughts
Evaluate the effectiveness from three aspects:
Execution time
Energy cost
Information latency
Karma: a. greedy dispatching b. continuous dispatching
Oracle: with foreknowledge of all activities, hence it can
give a lower bound of requirements.
Resilience to Failure
fail-->cannot update Datastore-->Dispatcher
detects that-->Scheduler re-arranges in next
allocation cycle.
Can be mitigated by reserving drones or
increasing the swarm size.
Adaptability
introduce a constant wind over the bottom
third of the field.
In that field, the round trip time reduces 32%
For same amount of work, 12% more drones.
Equally, 7% higher energy cost
Interesting feature: static VS active

Hive-drone paradigm can be used to continuously
measure time varying phenomena

Example: a chemical plume tracking application
Active
Sliding window
T:5m
Static

The current system is limited by the flight time of drones.
If the drones had a longer flight time, they can operate
different tasks per sortie which makes the system more
efficient.

The assumption in the design is that the drones cannot
communicate with each other. If the communication
could be achieved, the allocation policy might be
changed and information latency might also be reduced.

Inspired by idea of communication between two drones,
communication between two hives may improve the
coverage of application.
Download