Ready, Set, Go: Coalesced Offloading from Mobile Devices to the Cloud Liyao Xiang , Shiwen Ye, Yuan Feng, Baochun Li, Bo Li Department of Electrical and Computer Engineering University of Toronto May 1st, 2014
Remote execution Application Server External Storage Application Server 2
Code offloading Solver Solver Appli- Appli- cation cation Profiler Profiler Application Server Smartphone Cloud 3
Tail time phenomenon ‣ When multiple applications send their offloading requests without coordination, network interface enters at high-power state at arbitrary times. 4
Coalesced Offloading Offloading Requests Offloading Requests Coordination Service Coordination Service Solver Appli- Solver Solver cation App App Profiler Profiler Profiler Application Server Application Server Smartphone Cloud 5
Coalesced Offloading requests requests (a) Before bundling: of app 1 of app 2 Time t 1 t 2 t 3 t 4 t 5 t 6 t 7 Power State (b) After bundling: Time t 2( t 1 ' ) t 3 t 5( t 4 ' ) t 7( t 6 ' ) Power State 6
Problem Formulation ‣ Assume that M applications, generating requests at . The requests are granted at . a 1 , a 2 , ... g 1 , g 2 , ... High-power State Low-power State Power State Tail Time T latency(1) Time a 4( g 2) a 5( g 3) a 8( g 4) = t 2 a 9( g 5) a 1 a 2 a 3( g 1) = t 1 s 1 a 6 a 7 7
Problem Formulation X ‣ Energy cost function min { g j − g j − 1 , T } ' j X X ‣ Latency cost function ( g j − a i ) = ai s.t. j gj − 1 ≤ ai ≤ gj ‣ The joint optimization problem is as follows: X X X min f cost = min { g j − g j − 1 , T } + α ( g j − a i ) ai s.t. j j gj − 1 ≤ ai ≤ gj 8
How to solve the problem? 9
RSG Solutions ‣ Optimal offline algorithm: ‣ With the arrival time sequence a 1 , a 2 , . . . , a n known a priori . ‣ Online algorithms. ‣ Without a priori knowledge of the arrival time sequence. 10
RSG Offline Solution ‣ For request , a i ( min { a i − g prev , T } , if granted, f i cost = α ( g next − a i ) , if delayed. ‣ For Combinations of binary transmission sequence, 2 n we should: n X f i min f cost = cost i =1 ‣ The problem is transformed from continuous-time to discrete-time formulation. 11
What if we don’t know the entire input sequence? 12
Our Results ‣ Algorithm is 2-competitive . A 1 ‣ The competitive ratio between the expected cost incurred by and the A optimal cost is . e/ ( e − 1) ‣ RSG Online Algorithm have the optimal competitive ratio. 13
Performance Evaluation ‣ Measuring the Tail Time ( on iPhone 3GS, Bell Mobility 3G network ) ‣ Transmitting successive packets of equal size with constant transmission intervals. ‣ Model-driven Simulations ‣ Simulating the timing of multiple offloading requests from several simultaneously running applications. ‣ Real-world Experiments 14
Experiment Results Naive Unit Time Energy Cost (mAh) 0.08 Deterministic Randomized Energy consumption with 0.07 Offline 0.06 different types of requests 0.05 0.04 0.03 0.02 Random Bursty Stable Naive Unit Time Energy Cost (mAh) Offline 0.09 Randomized 0.08 Deterministic Energy consumption with varying 0.07 0.06 alpha 0.05 0.04 0.03 0.02 15 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 α
Experiment Results Rubik Solver Email Chat Real requests on mobile device Request Grants w/o RSG solutions Naive Online Raw Battery Voltage (mV) Time (s) 4140 4120 4100 4080 Battery Voltage Change on 4060 4040 50 100 150 200 250 mobile device w/o RSG Raw Battery Voltage (mV) 4140 solutions 4120 4100 4080 4060 4040 50 100 150 200 250 16 Time (s)
Conclusions ‣ By bundling the offloading requests of multiple applications, we achieve greater energy savings while maintaining satisfactory performance. ‣ The RSG online algorithm achieves the best possible competitive ratio. 17
Thank you. 18
Recommend
More recommend
