Dynamic Mobile Device Integration for Efficient Cross-device Resource Sharing

Transcription

1 : Dynamic Mobile Device Integration for Efficient Cross-device Resource Sharing Dongju Chae 1, Joonsung Kim 2, Gwangmu Lee 2, Hanjun Kim 1, Kyung-Ah Chang 3, Hyogun Lee 3, and Jangwoo Kim 2 1 Dept. of Computer Science and Engineering, POSTECH 2 Dept. of Electrical and Computer Engineering, Seoul National University 3 Samsung Electronics

2 Increasing Number of IoT Devices 2/35

3 Sharing Multi-device Enables Many Services I want to build my home theater Audio Smart Monitoring System? Custom Home Theater Storage Remote Gaming System?, and various custom services. Audio 3/35

4 Index Limitations of existing schemes I/O request forwarding Manual programming : Efficient dynamic resource sharing Evaluation Conclusion 4/35

5 Running Example: Home Theater Program open(storage); load(data); decode(data); render(data); 5/35

6 (1) I/O Request Forwarding - with the example of home theater Phone TV Program open(storage); load(data); decode(data); render(data); Decoded data I/O abstraction layer High Traffic! I/O Resources data Storage I/O Resources Storage 6/35

7 (1) I/O Request Forwarding - with the example of home theater Good load(data); decode(data); render(data); Storage Phone TV Program I/O open(storage); abstraction layer for transparency Easy programming environment Bad Decoded data I/O abstraction layer High Traffic! High network traffic duei/o to Resources unoptimizedi/o datapath Resources data Poor performance Storage 7/35

8 Sender Receiver (2) Manual Programming - with the example of home theater Phone Program TV Program open(storage); load(data); Low Traffic! recv(data); send(data); decode(data); I/O Interface Decoded data I/O Interface render(data); I/O Resources data Storage I/O Resources Storage 8/35

9 Sender Receiver (2) Manual Programming - with the example of home theater Good open(storage); open(storage); load(data); load(data); send(data); decode(data); render(data); Phone Program I/O Interface Phone Program Device-aware task partitioning Low Good performance Traffic! Bad TV Program Decoded data I/O I/O abstraction Interface layer TV recv(data); decode(data); render(data); Hand-tuned multi-device application data High programming effort I/O ResourcesI/O Resources I/O ResourcesI/O Resources Storage Storage Storage Storage 9/35

10 Programmability Design Goals Request Forwarding High Performance & Programmability Manual Programming Performance 10/35

11 Index Limitations of existing schemes : Efficient dynamic resource sharing Key ideas Architecture & Implementation Evaluation Conclusion 11/35

12 Key Ideas of 1) Transparent & Wide Resource Integration Kernel-level resource integration Transparent integration (easy programmability) Sharing multiple types of resources (i.e., CPU, Memory, I/O) Wide resource coverage 2) Resource-aware Dynamic Task Redistribution Contention detection based on resource usage Performance estimation for migration scenarios Optimized performance 12/35

13 Key Ideas of (1/2) 1) Transparent & Wide Resource Integration I want to use remote screen on TV! I/O datapath Local I/O Remote I/O User Kernel I/O Resource I/O Resource Phone TV 13/35

14 Key Ideas of (2/2) 2) Resource-aware Dynamic Task Redistribution I/O datapath Local I/O Remote I/O User High Traffic! Kernel I/O Resource I/O Resource Phone TV 14/35

15 Key Ideas of (2/2) 2) Resource-aware Dynamic Task Redistribution Low Traffic! User I/O datapath Local I/O Remote I/O Kernel I/O Resource I/O Resource Phone TV 15/35

16 Architecture User Migration Selector Migration Selector Kernel Contention Detector Contention Detector (1) Resource Integrator Select Resource Integrator Storage Local Device (Master) Remote Device 16/35

17 Architecture User Migration Selector Migration Selector Kernel (2) Contention Detector Contention Detector Resource Integrator Resource Integrator Storage Local Device (Master) Remote Device 17/35

18 Architecture User Kernel (3) Migration Selector Migration Selector Contention Detector Contention Detector Resource Integrator Resource Integrator Storage Local Device (Master) Remote Device 18/35

19 (1) Resource Integrator (Memory, CPU, I/O) DSM Memory integration Distributed shared memory (DSM) Three perf. optimizations Lazy release consistency model Page-level coherency block Memory prefetching Device Device 19/35

20 (1) Resource Integrator (Memory, CPU, I/O) Migration CPU integration Thread migration DSM Optimizations Thread group granularity Clone-based migration Transparent live migration Device Device 20/35

21 (1) Resource Integrator (Memory, CPU, I/O) Migration I/O resource integration Request forwarding (e.g., device file) DSM /dev/xx Device File /dev/xx Optimizations Data compression Platform-assisted handling Device Resource XX Device 21/35

22 2) Contention Detector Calculate the slowdown of each thread Collect per-thread resource usage (e.g., CPU, network, ) Measure a stall time due to resource access File user Comp kernel Low (fast) High Storage Task Latency Phone TV 22/35

23 2) Contention Detector Calculate the slowdown of each thread Collect per-thread resource usage (e.g., CPU, network, ) Measure a stall time due to resource access File user Comp Contention Low (fast) Network High Storage kernel Task Latency Phone TV 23/35

24 3) Migration Selector Estimate the performance tradeoff after migration Utilize intra-/inter-device information to decide migration targets 1. Intra-device communication 2. Inter-device bandwidth & latency Group 1 Group 2 1MB CPU 2MB Storage Player 5MB 5MB 3. Remote resource status Phone TV 24/35

25 3) Migration Selector Estimate the performance tradeoff after migration Utilize intra-/inter-device information to decide migration targets Calculate the tradeoffs of all possible migration scenarios High traffic! High traffic! High mig. overhead! Scenario #1 Scenario #2 Scenario #3 Low traffic! Scenario #4 Select Target Storage Storage Storage Storage

26 Index Limitations of existing schemes : Efficient dynamic resource sharing Evaluation Scenario #1: Home theater Scenario #2: Smart monitoring Conclusion 26/35

27 Evaluation Setup: Home theater Master Device (Local) Home Theater Large TV (Remote) HQ Speaker (Remote) Android-based smartphones (5.1.1) - Nexus 5 (master) - Nexus 4 (connected to a speaker) Tizen-based smart TV (2.3) Storage Audio Phone (Master) Speaker Audio TV 27/35

28 Frames per second (FPS) Performance Timeline - Resource reconfiguration (local screen remote large screen) 30 Local Remote Large Select a remote large screen Start migration Restore the target performance 6 0 Detect contention & Start analysis Time (sec) /35

29 Frames Per Second Throughput Improvement - Measured FPS for the home theater scenario Achieve (or closely) the target performance goal 30 Ideal (24 FPS) Request Forwarding 240p 360p 480p 720p Video quality (HD) 4x higher 8.3x higher 1080p (Full HD) 29/35

30 Avg. Stall Per Frame (ms) Minimized Network Stall Time - Per-frame latency analysis for each datapath Reduce the exposed network traffic Storage Phone TV Request Forwarding (RF) Storage Phone TV (DM) 50 0 RF DM RF DM RF DM RF DM 240p 480p 720p (HD) Video quality p (Full HD) 30/35

31 Evaluation Setup: Smart Monitoring Smart Monitoring Master Device (Local) Selector Processing Phone Processing Camera Camera Device (Remote) Android-based smartphones (5.1.1) - Nexus 5 (master) - Nexus 5 (camera) x 3 Camera Camera Phone (Master) Camera Phone 31/35

32 Frames Per Second Throughput Improvement - As the camera preview resolution increases Achieve higher throughput than Request Forwarding Ideal (15 FPS) Request Forwarding 176x x x x x480 Preview resolution (pixels) 3x higher 32/35

33 Avg. Frame Latency (ms) Computation Bottleneck - Per-frame detection latency Potential to achieve higher throughput with faster CPUs 400 Network Computation RF DM RF DM RF DM RF DM RF DM 176x x x x x480 Preview resolution (pixels) 33/35

34 Conclusion : Efficient cross-device resource sharing Transparent resource integration for diverse resources Resource-aware dynamic task redistribution Implementation on top of Android/Tizen devices Achieve target performance for multi-device services e.g., 8.2 FPS 24 FPS on the home theater scenario 34/35

35 : Dynamic Mobile Device Integration for Efficient Cross-device Resource Sharing Thank You! Dongju Chae Department of Computer Science & Engineering Pohang University of Science and Technology (POSTECH) 35/35