1. 1 Mobile Software Development Framework: Adaptive Mobile Applications 10/23/2012 Y. Richard Yang

2. 2 Outline r Admin r Mobile cloud services m Push notification service m Track service m Storage service r Adaptive mobile applications

3. 3 Admin. r HW3 r Project ideas

4. Recap: Example Mobile Cloud Services r Push notification service r Location based services, e.g., m Track service (supporting location based services) r Storage services, e.g., m iCloud, Google Drive, Dropbox r Proxy services, e.g., m Kindle Split Browser m Recognition/synthesis services 4

5. Example: Push Notification Service r A single persistent connection between Push Notification Service (PNS) and each mobile device r Authorization m App developer registers project with PNS m App on device registers with PNS and forwards registration to PNS m App server sends to registered apps on devices r Scalability, fault tolerance, generality m Send only notification, not data m For reliability, use soft state, not hard state 5

6. Example: Track Service  TC JoinTrackCollections (TC tCs[], bool removeDuplicates)  TC SortTracks (TC tC, SortAttribute attr)  TC TakeTracks(TC tC, int count)  TC GetSimilarTracks (TC tC, Track refTrack, float simThreshold)  TC GetPassByTracks (TC tC, Area[] areas)  TC GetCommonSegments(TC tC, float freqThreshold) Manipulation Creation  TC MakeCollection(GroupCriteria criteria, bool removeDuplicates) 6

7. API Usage: Ride-Sharing Application // get user’s most popular track in the morning TC myTC = MakeCollection(“name = Maya”, [0800 1000], true); TC myPopTC = SortTracks(myTC, FREQ); Track track = GetTracks(myPopTC, 0, 1); // find tracks of all fellow employees TC msTC = MakeCollection(“name.Employer = MS”, [0800 1000], true); // pick tracks from the community most similar to user’s popular track TC similarTC = GetSimilarTracks(msTC, track, 0.8); Track[] similarTracks = GetTracks(similarTC, 0, 20); // Verify if each track is frequently traveled by its respective owner User[] result = FindOwnersOfFrequentTracks(similarTracks);

8. Outline r Admin and recap r Mobile Cloud Services m Push notification service m Track service m Storage service 8

9. Storage/Sync Service r Store content in cloud to be accessible by multiple selected devices r Multiple deployed services, e.g., m iCloud m Google Drive m DropBox 9

10. Storage/Sync Service Example: iCloud r Backend m Hosted by Windows Azure and Amazon AWS m Uses HTTPS connect to servers r More details: see iCloud sessions m https://developer.apple.com/videos/wwdc/2012/?id =209 10

11. 11 Key Design Points for Mobile Storage r Data models m what data structure to put on mobile storage r Read miss m stalls progress (user has to wait for data) r Synchronization/consistency m user may see outdated data m user modification may generate conflicts r Good topic for investigation/design projects m See backup slides for some pointers

12. Outline r Admin and recap r Adaptive mobile applications 12

13. Adaptive Mobile Applications r An adaptive mobile application adjusts the amounts as well as the locations of resources that it consumes r The objective typically is to m Improve responsiveness m Reduce energy consumption at device 13

14. Amazon Silk Split-Browser r Dynamically split browsing work between local device and Amazon cloud 14 http://www.extremetech.com/mobile/97587-amazon-silk-bridging-the-gap-between-desktop-and-tablet-web-browsers

15. 15 Adaptive Mobile App: Big Picture device in-net proxy service server On-device app/sys adaptation Service partition Device-aware service partition, delivery

16. Adaptive Mobile Applications: Overview r This is a large topic r We will look at an example on each entity 16 device in-net proxy service server On-device app/sys adaptation Service partition Device-aware service partition, delivery

17. Outline r Admin and recap r Adaptive mobile applications m Device adaptation 17

18. Discussion r What resource parameters/aspects of a mobile device can be controlled? 18 http://www.wired.com/gadgetlab/2 012/09/iphone5-spec-showdown/

19. Example: iphone 5 19 http://www.anandtech.com/show/6324/the-iphone-5-performance-preview iPhone 5 controls the CPU frequency idle downloading

20. 20 Iphone5 A6 Processor http://www.zdnet.com/inside-apples-a6- processor-7000004786/

21. 21 CPU Power Model r 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 r When the supply voltage V is lower, charging/discharging time is longer; thus frequency should be lower => P ~ O(V 3 )

22. 22 CPU Power Model Question: r Suppose P ~ V 3. Do we save energy by reducing freq/vol to 1/3, but finish the same job? r What if linear: P ~ V? throughput

23. 23 Dynamic Voltage Scaling r For convex power consumption rate, to minimize power for finishing a job, one may lower the freq/vol to close to 0 r Constraint m bound response time (e.g., deadline of multimedia playback)

24. 24 Example Architecture: GraceOS CPU monitoring scheduling speed scaling demand distribution scheduler speed adaptor profiler multimedia applications requirements time constraint

25. 25 Demand Prediction  Online profiling and estimation: count number of cycles used by each job b 1 b 2 C min =b 0 b r =C max 1 b r-1 cumulative probability CDF F(x) = P [X  x]

26. 26 Demand Stability Demand distribution is stable or changes slowly

27. 27 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 ]   b 1 b 2 b 0 brbr 1 b r-1 cumulative probability F(x)F(x)  C

28. 28 How Fast to Run the CPU?  Goal: provides C i cycles during a time duration of P i  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.

29. 29 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

30. 30 Stochastic DVS For each job 1.find speed S x for each unit allocated cycle x b 1 b 2 C min =b 0 b r =C max 1 b r-1 cumulative probability CDF F(x) = P [X  x]

31. 31 Stochastic DVS For each job 1.find speed S x for each allocated cycle x  time is 1/S x and energy is (1 - F(x))S 3 x such that

32. 32 Example Speed Schedule 100 MHz200 MHz400 MHz cycle: speed: Job 1 2.5 x10 6 cycles speed (MHz) 100 400 200 Job 2 1.2 x10 6 cycles Observation: speed up the processor with increasing clock cycles

33. 33 DVS speed A1 B1 A1 execution speed up within job context switch 1.Store speed for switched-out 2.New speed for switched-in switch back A2

34. 34 Implementation Hardware: HP N5470 laptop –Athlon CPU (300, 500, 600, 700, 800, 1000MHz)  round speed schedule to upper bound process control block standard Linux scheduler DVS modules PowerNow speed scaling Soft real-time scheduling system call Extension to Linux kernel 2.4.18 716 lines of C code

35. 35 Reduces power consumption However, limited due to few speed options Evaluation: Normalized Energy

36. 36 Odyssey: An Example Client Architecture r Application indicates resource capabilities in its request to service r Operating system maintains/monitors available resources m no need to have each application re-implement the monitoring r An application registers a resource descriptor and an upcall event handler with the OS r OS notifies the application upon detecting resource changes r Application adjusts requests to the server

37. Outline r Admin and recap r Adaptive mobile applications m Device adaptation m Server adaptation 37

38. 38 Server Adaptation r It is a good practice to design your app so that m Client notifies the server about its capability m Server adapts according to client capability

39. 39 Example: Content Distribution Server r Server has many dimensions to adapt image/video fidelity: m frame rate (for video) m image size m quality of image r Usage: e.g., data acceleration offered by many carriers

40. 40 Frame Encoding: Block Transform Encoding DCT

41. 41 Discrete Cosine Transform F[u,v] = 4C(u)C(v) n2n2  n-1 j=0 k=0 f(j,k) cos (2j+1)u  2n (2k+1)v  2n cos where C(w) = 1  2 1 for w=0 for w=1,2,…,n-1

42. 42 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] Q: If you want to change encoding rate, what may you do? more detail

43. 43 Frame Encoding: Block Transform Encoding DCT Zig-zag Quantize Run-length Code Huffman Code 011010001011101...

44. 44 Example: MPEG Block Encoding original image DCT DC component AC components Quantize zigzag run-length and Huffman encoding of the stream 10011011100011... coded bitstream < 10 bits (0.55 bits/pixel)

45. 45 Examples Uncompressed (262 KB) Compressed (22 KB, 12:1) Compressed (6 KB, 43:1)

46. Outline r Admin and recap r Adaptive mobile applications m Client adaptation m Server adaptation m Proxy and job partition 46

47. Discussion r When may a proxy help? 47 device in-net proxy in-net server

48. Example: Round-Trip-Time Effects 48 device server

49. Example: Round-Trip-Time Effects 49 device server

50. Measured Benefit of Reduced Latency 50

51. Example System: MAUI 51 Maui server Smartphone Application Client Proxy Profiler Solver Maui Runtime Server Proxy Profiler Solver Maui Runtime Application RPC Maui Controller

52. How Does a Programmer Use MAUI? m Add.NET attributes to indicate “remoteable”

53. Maui server Smartphone Application Client Proxy Profiler Solver Maui Runtime Server Proxy Profiler Solver Maui Runtime Application RPC Maui Controller Intercepts Application Calls Synchronizes State Chooses local or remote Handles Results Provides runtime information

54. Profiler Callgraph Execution Time State size Network Latency Network Bandwidth Device Profile CPU Cycles Network Power Cost Network Delay Computational Delay Computational Power Cost Computational Delay Annotated Callgraph

55. B 900 mJ 15ms C 5000 mJ 3000 ms 1000mJ 10ms 25000 mJ 5ms D 15000 mJ 12000 ms 10000 mJ 5ms A Computation energy and delay for execution Energy and delay for state transfer A sample callgraph

56. B C D A

57. FindMatch 900 mJ InitializeFace Recognizer 5000 mJ 1000mJ 25000 mJ DetectAndExtract Faces 15000 mJ 10000 mJ User Interface

58. FindMatch 900 mJ InitializeFace Recognizer 5000 mJ 1000mJ 25000 mJ DetectAndExtract Faces 15000 mJ 10000 mJ User Interface Cheaper to do local Compare cost of state transfer and local comp. Pick lower cost

59. FindMatch 900 mJ InitializeFace Recognizer 5000 mJ 1000mJ 25000 mJ DetectAndExtract Faces 15000 mJ 10000 mJ User Interface Cheaper to do local Compare cost of state transfer and local comp. Pick lower cost

60. FindMatch InitializeFace Recognizer 1000mJ DetectAndExtract Faces User Interface 20900mJ Cheaper to offload

61. MAUI Implementation r Platform m Windows Mobile 6.5 m.NET Framework 3.5 m HTC Fuze Smartphone m Monsoon power monitor r Applications m Chess m Face Recognition m Arcade Game m Voice-based translator

62. 4x savings on 3G 8x improvement on Wi-Fi Face Recognizer

63. 9x improvement on wifi; 5x on 3G Face Recognizer

64. Up to 40% energy savings on Wi-Fi Solver would decide not to offload Arcade Game

65. Backup Slides

66. 66 Mobile Storage r Read miss m explicit user file selection m automatic hoarding/prediction, e.g., CODA, SEER r Synchronization/consistency m keep modification logs and develop merge tools, e.g., Bayou m efficient file comparisons and merging, e.g., rsync, LBFS

67. 67  One approach is to keep modifications as logs  a user sends the logs to the servers to update  If the storage of a client is limited, it may not be able to save logs  then upon reconnection, the cache manager needs to find the difference between the stored file and its local cached copy  same problem exists for the rsync tool !  Question: how to efficiently compare the differences of two remote files (when the network connection is slow)?

68. 68  Break Files into chunks and transfer only modified chunks  Fixed chunk size does not work well  why?

69. 69  Compute hash value of every 48 byte block  if the hash value equals to a magic value, it is a chunk boundary

70. 70 Bayou: Managing Update Conflicts r Basic idea: application specific conflict detection and update r Two mechanisms for automatic conflict detection and resolution m dependency check m merge procedure http://zoo.cs.yale.edu/classes/cs422/2011/bib/terry95managing.p df

71. 71

72. 72