Adaptive Mobile Applications 4/21/2009 Richard Yang

Admin. Project

Challenges in Developing Mobile, Wireless Applications Need to adapt to heterogeneous, dynamic device/network status Need to optimize resource usage given that resources are often limited

Motivation for Service Adaptation Service delivered should depend on device status (e.g., display capability, input method, memory size, remaining battery level) network status in-building campus-area packet radio metro-area regional-area

Adaptation: Who and How? request client service result Adaptation need support of the client, the server, and/or a third party (proxy or gateway) Server and third party adaptation is also called content adaptation

Adaptation: Client Responsibility Client informs server its device capability and connection status Client makes best use of available resources and available result data

Content Adaptation Server/third party adapts service/content according to client capability

Client Adaptation

Client Capability Negotiation and Monitoring Rudimentary capability already exists, but far from complete

Two Examples of Client Makes Best Use of Resources/Results Adaptive CPU scheduling Adaptive playout

Example 1: Adaptive CPU Power Management

Big Picture: Power Management

Power Model

Some Techniques of Power Management

Dynamic Voltage Scaling Dynamic Voltage Scaling (DVS) allows voltage to be dynamically adjusted DVS-enabled processors, e.g., StrongARM SA-2 (500mW at 600MHz; 40mW at 150MHz) AMD (PowerNow!) http://www.amd.com/us-en/assets/content_type/DownloadableAssets/Power_Now2.pdf Intel (SpeedStep, XScale) Texas Instrument Transmata

CPU Power Consumption When the supply voltage V is lower, charging/discharging time is longer; thus frequency should be lower

CPU Power Consumption Model The power consumption rate P of a CMOS processor satisfies where k is a constant, C the capacitance of the circuit, f the CPU frequency, and V the voltage -> P ~ O(V3)

Voltage Scaling throughput Discussion: what voltage to operate on?

Dynamic Voltage Scaling Basic idea: determining voltage according to program response time requirement For normal applications, give reasonable response time For multimedia applications, use the deadline to determine voltage

multimedia applications Architecture multimedia applications monitoring requirements scheduling scheduler demand distribution profiler time constraint speed adaptor speed scaling CPU

Demand Prediction Online profiling and estimation: count number of cycles used by each job 1 CDF F(x) = P [X  x] cumulative probability Cmin=b0 b1 b2 br-1 br=Cmax

Demand distribution is stable or changes slowly Observations Demand distribution is stable or changes slowly

CPU Resource Allocation How many cycles to allocate to a multimedia job? Application should meet  percent of deadlines  each job meets deadline with probability   allocate C cycles, such that F (C ) =P [X  C ]   b1 b2 b0 br 1 br-1 cumulative probability F(x)  C

How Fast to Run the CPU? Assume the strategy is to run job i at a fix (also called uniform) speed Si Assume it needs Ci cycles during a time duration of Pi Fact: since power is a convex function of frequency, if a job needs C cycles in a period P, then the optimal frequency is C/P, namely the lowest constant frequency.

Why Not Uniform Speed? Intuitively, uniform speed achieves - minimum energy if use the allocated exactly However, jobs use cycles statistically - often complete before using up the allocated - potential to save more energy  stochastic DVS

Stochastic DVS For each job such that find speed Sx for each allocated cycle x time is 1/Sx and energy is (1 - F(x))S3x such that

Observation: speed up the processor with increasing clock cycles Example Speed Schedule cycle: speed: 100 MHz 200 MHz 400 MHz Job 1 2.5x106 cycles speed (MHz) 100 400 200 Job 2 1.2x106 cycles Observation: speed up the processor with increasing clock cycles

DVS A1 B1 A1 A1 A2 context switch Store speed for switched-out New speed for switched-in speed up within job switch back A1 B1 A1 speed A1 A2 execution

Extension to Linux kernel 2.4.18 Implementation Hardware: HP N5470 laptop Athlon CPU (300, 500, 600, 700, 800, 1000MHz) round speed schedule to upper bound system call process control block DVS modules PowerNow speed scaling Soft real-time scheduling standard Linux scheduler Extension to Linux kernel 2.4.18 716 lines of C code

Evaluation: Normalized Energy Reduces power consumption However, limited due to few speed options

Example 2: Deal with Variable Delay

Example: Deal with Variable Delay Consider multimedia applications Packets arrive with variable delay (why?) Variable delay does not work well for multimedia application

Client Playout Buffer Basic idea: delay playing out packets to accommodate delay jitter Question: how to determine the playout delay? packets number time packets generated p p' delay 1 r received

Odyssey: An Example Client Architecture Application indicates resource capabilities in its request to service Operating system maintains/monitors available resources no need to have each application re-implement the monitoring An application registers a resource descriptor and an upcall event handler with the OS OS notifies the application upon detecting resource changes Application adjusts requests to the server

Client System Model

Content Adaptation

Objective: automating adaptation Content Adaptation: Examples Objective: automating adaptation

Example: Adapting Audio Content 2 3 Send a lower resolution stream as the redundant information, e.g. nominal stream at 64 kbps and redundant stream at 13 kbps (such as GSM)

Example: Adapting Fidelity of Video/Image Content? Potentially many dimensions frame rate (for video) image size quality of image Usage: e.g., data acceleration offered by many carriers

Frame Encoding: Block Transform Encoding DCT Zig-zag Quantize Huffman Code Run-length Code 011010001011101...

Discrete Cosine Transform 4C(u)C(v) n-1 n-1 (2j+1)up (2k+1)vp å å f(j,k) cos F[u,v] = cos n2 2n 2n j=0 k=0 where 1 for w=0 Ö 2 C(w) = for w=1,2,…,n-1 1 DCT is better at reducing redundancy than Discrete Fourier Transform but it is more computationally expensive

Basis Functions of DCT An image is a superposition of basis functions DCT computes the contribution of each basis function - F[u,v]: for the basis function at position [u, v] more detail

Example: MPEG Block Encoding DC component Quantize DCT original image AC components zigzag run-length and Huffman encoding of the stream 10011011100011... coded bitstream < 10 bits (0.55 bits/pixel) Discussion: how to generate different encoding rates?

Examples Uncompressed (262 KB) Compressed (22 KB, 12:1)