Non-Cooperative Computation
Offloading in Mobile Cloud Computing
Valeria Cardellini
University of Roma Tor Vergata, Italy
Joint work with: V. De Nitto Personè, V. Di Valerio, F. Facchinei, V. Grassi, F. Lo Presti, V. Piccialli
Joint CLEEN and ACROSS Workshop on Cloud Technology and Energy Efficiency in Mobile Communications
Computation offloading
n
Mobile devices have
limited
computing and
processing
capabilities
¨ Battery life!
n
“Old” idea:
offload computation
to some
surrogate server (cyber foraging)
n
Mobile-cloud convergence
n
Increasing number of latency-sensitive,
compute-intensive and memory-intensive
applications on mobile devices
¨ Wearable cognitive assistance, edge analytics in IoT,
automotive environments, …
Cloud-‐based augmenta0on of mobile
devices
n
Resource-‐limited mobile devices can overcome their
limita0ons by offloading opera0ons to more
powerful servers
¨ Either proximate, i.e. single hop (cloudlet) or mul0-‐hop
(cloud data center)
Mobile devices Cloudlet Distant cloud servers on Internet Wireless AP
3 Non-Cooperative Computation Offloading in Mobile Cloud Computing @EUCNC'15 Workshop 1, June 29 2015
Cloudlet
n
Self-‐managed data center with soC state and located
at the network edge
¨ Consistent with Fog compu0ng (“Fog is the cloud close to
the ground”) and micro cloud concepts
n
Example of cloudlet infrastructure and soCware
Nokia LTE base sta0on with cloudlet-‐like hos0ng infrastructure
Cloudlet-‐specific extensions to OpenStack (by Satya at CMU)
Three-‐0er architecture
n
Hybrid three-‐0er
architecture for mobile-‐cloud
convergence
¨ Local 0er: mobile nodes
¨ Middle 0er: energy-‐unlimited but resource-‐limited nearby
cloudlets within one wireless hop of mobile devices
¨ Remote 0er: resourceful distant cloud servers
n
Where to offload?
¨ How to get the benefits of external servers for offloading
while limi0ng latency issues of distant servers and considering energy constraints of mobile devices
Our proposal
n
Non-‐coopera0ve
game theore0c
approach to
analyze and manage the computa0on offloading
strategy
¨ Unmanaged usage scenario: no central authority
¨ Mul0ple non-‐coopera0ve mobile devices share the
cloudlet limited compu0ng resources
¨ Mobile devices can send their computa0ons to any of the
three 0ers
V. Cardellini, V. De NiTo Personè, V. Di Valerio, F. Facchinei, V. Grassi, F. Lo Pres0, V. Picciali, “A game-‐theore0c approach to computa0on offloading in mobile cloud compu0ng”, Mathema5cal Programming, available online, Apr. 2015.
Our goal
n
Determine
whether and where to offload
a task
based on the impact this has on the user experience,
expressed through
QoS measures
System model
nStandard queueing
theory for system model
¨ To capture the effects of
computa0on offloading (i.e., task conten0on on shared resources) on users’ perceived
performance
¨ Focus on task
response 0me
Problem formula0on
nTo determine the
op0mal strategy, each
user needs to solve:
n
Problem formula0on:
generalized Nash equilibrium
problem
(GNEP)
¨ Both the players objec0ve func0ons and the strategy sets
depend on the other players strategies
GNEP proper0es and distributed algorithm
nGame can be solved by finding a solu0on to a
suitable
Varia0onal Inequality
¨ We show equilibrium existence and monotonicity
proper0es of the GNEP
n
Distributed algorithm
for compu0ng an
equilibrium strategy for each mobile user
n
Its design required us:
¨ To analyze in-‐depth the underlying equilibrium
problem
¨ To apply some very recent algorithmic developments n
Appealing for real-‐world implementa0on
Distributed algorithm
nSynchronous and itera0ve
Numerical results
Non-Cooperative Computation Offloading in Mobile Cloud Computing @EUCNC'15 Workshop 1, June 29 2015 12
cloudlet
mobile
cloudlet
cloud
Numerical results
Non-Cooperative Computation Offloading in Mobile Cloud Computing @EUCNC'15 Workshop 1, June 29 2015 13
n
Price of anarchy
cloudletcloud cloudlet
Prototype architecture
Non-Cooperative Computation Offloading in Mobile Cloud Computing @EUCNC'15 Workshop 1, June 29 2015 14
n
Working prototype
based on offloading middleware
Prototype addi0onal features
n
Adapta0on capabili0es based on MAPE cycle
n
Discovery and deployment of offloading service
Prototype results: face recogni0on
Non-Cooperative Computation Offloading in Mobile Cloud Computing @EUCNC'15 Workshop 1, June 29 2015 16
Conclusions
n
Non-‐coopera0ve usage scenario for computa0on
offloading in a three-‐0er architecture
¨ Users try to take advantage selfishly of the available
cloudlet resources
¨ Solu0on with solid theore0cal founda0on and appealing
for real-‐world implementa0on
Future work
n
Possible extension of our proposal
¨ Remove limita0on on user generated load (cannot exceed
the mobile device capacity)
¨ Consider mul0ple class model of tasks launched by users
¨ Include monetary cost n
More general challenges
¨ Use cloudlets to improve reliability
¨ Consider mul0ple (possibly coopera0ve) cloudlets
That’s All Folks!
19