Energy-Delay Tradeoffs in Smartphone Applications Moo-Ryong Ra Jeongyeup Paek, Abhishek B. Sharma Ramesh Govindan, Martin H. Krieger, Michael J. Neely Mobisys 2010
Introduction
Battery lifetime Urban Tomography system users reported that battery lifetime is a critical usability issue
Energy-delay tradeoff
Algorithm They design a control algorithm, called SALSA (Stable and Adaptive Link Selection Algorithm): use the Lyapunov optimization framework - minimizes the total energy expenditure subject to keeping the average queue length finite This algorithm considers link selection problem
Problem statement, model and objective
Notations A[t]: the size of video data in bits P[t]: the power consumption , μ[t]: the amount of data transferred U[t]: queue backlog L[t]: set of links visible to a smartphone [t]: the quality of the wireless link I[t]: indicator random variable 0, 1 I[t] == 1 >> smartphone decides to transmit data I[t] == 0 >> otherwise
Model and objective μ[t] ≜ C(I[t], l, [t], U[t], P[t]) U[t+1] = U[t] - μ[t] + A[t] Stability Minimizes the time average transmit power expenditure
The link selection algorithm
SALSA’s control decision SALSA decides, every timeslot t, whether to transmit data from its queue, and which of its available links to use The performance of this algorithm critically depends upon the choice of V SALSA’s control decision using the Lyapunov optimization
Constraints Power consumption satisfying: Queue backlog satisfying: Trade-off between power consumption and delay depend on the parameter V: [O(1/V), O(V)] P* is a theoretical lower bound on the time average power consumption B is an upper bound on the sum of the variances of A[t] and μ[t]
Choosing a good V V controls the energy-delay tradeoff (α is the slope of ) Adapt V to the instantaneous delay D[t] denotes the instantaneous delay in data transfer SALSA computes B based on all the A[t] and μ[t] values observed over some large time window It updates its value whenever the estimate for B is updated Instead of using a different parameter, they chose to use α in order to have only one free parameter in SALSA
Evaluation
Overview They use trace-driven simulation - arrival traces: derived from users of their urban tomography system in real-world settings - link availability traces: generated empirically by carrying a smartphone on a walk across different environments They compare SALSA against two baseline algorithms: - minimize delay and always uses WiFi
Arrival Patterns They use a total of 42 arrival patterns consisting of a total of 935 videos
CDF link availability with failure probability CDF of the average transfer rate per 20-second window USC campus A large shopping mall near Los Angeles (Glendale Galleria) Los Angeles International Airport (LAX)
Comparison Minimum-delay algorithm WiFi-only algorithm Static-delay algorithm Know-WiFi algorithm Minimum-delay algorithm: always transfers data when an AP is available (High energy) WiFi-only algorithm: uses only WiFi APs (Unbounded delay) Static-delay algorithm: it has not seen any WiFi AP in the past T timeslots, it uses the first link that becomes available (Not take link quality into account) Know-WiFi algorithm: assumes information about the availability of WiFi APs in the future (Not consider queue backlog)
Performance metrics The average energy consumed per byte - The average delay per byte - Dispersion
Minimum-delay vs WiFi-only vs SALSA
SALSA’s performance
Comparison with threshold-based algorithms 관점이 달라서 관점을 하나로 통일 >> the most aggressive value, the least aggressive value
Sensitivity to the scanning interval Four additional scanning intervals: 60s, 120s, 180s, and 240s The sweet spot for the scanning interval appears to be 60 seconds They simulated HD traffic in a single collision domain under varying densities and different bitrates Back2F provides gains are in the range of 15% to 30%
Experimental results
Environment They implement SALSA in a video transfer application developed in Symbian C++ for the Nokia N95 smartphone One volunteer carried five phones each configured with different values of α, and conducted 5 walks
Experimental result At the USC Campus At Shopping Mall
Summary
Summary Adaptive algorithm for energy/delay trade off - Extensive evaluation with real world scenarios - Validation with real implementation - Provable performance bound