HuaJun Hong June 2014 1 Publications H Hong D Chen C Huang K Chen and C Hsu QoS aware virtual machine placement for cloud games in Proc of ACM Annual Workshop on Network and Systems Support for Games ID: 1001812
Download Presentation The PPT/PDF document "Placing virtual machines to optimize clo..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1. Placing virtual machines to optimize cloud gaming experienceHua-Jun, HongJune, 20141
2. PublicationsH. Hong, D. Chen, C. Huang, K. Chen, and C. Hsu, “QoS-aware virtual machine placement for cloud games,” in Proc. of ACM Annual Workshop on Network and Systems Support for Games (NetGames), Denver, CO, December 2013.H. Hong, D. Chen, C. Huang, K. Chen, and C. Hsu, “Placing virtual machines to optimize cloud gaming experience,” submitted to IEEE Transactions on Cloud Computing, April 2014.H. Hong, C. Lee, K. Chen, C. Huang, and C. Hsu, “GPU consolidation for cloud games: Are we there yet?,” submitted to a double-blinded conference, under review, April 2014.2
3. outlineIntroductionProblem FormulationQDH AlgorithmTestbedTrace-Driven SimulationsMeasurement Study of Modern GPU3
4. The new approach to play gamesCloud gaming systemPlay any game at anytime, anywhere on any device!The increasing market!8 Billion USD!20174
5. Financial difficulty of providersCloud gaming providersGaikaiUbitusOnlive – runs into the financial difficulty…The financial problem may cause by…Network latencyResource allocation of each VMNon-mature GPU virtualization…5
6. Problem statementDiverse gaming hardware requirements may lead to wasted hardware resourcesConsolidating different games results in different profits and gaming qualityHence, we propose a VM placement policy to maximize the profits while achieve just-good-enough QoEAlso, we conduct a measurement study to make sure that if the modern GPU is powerful enough for the cloud gaming system6
7. GoalsFind the best tradeoff between gaming Quality-of-Experience and profitsAnswer the question that “Are modern GPUs ready for cloud gaming?”7
8. outlineIntroductionProblem FormulationQDH AlgorithmTestbedTrace-Driven SimulationsMeasurement Study of Modern GPU8
9. NotationsNetwork Latency: Frame Per Second: Processing Delay: CPU Utilization:GPU Utilization:Hourly fee:Operational Cost: Memory of Server:Uplink of Datacenter: 9
10. ModelsCPU utilization, GPU utilization, frame rate, and processing delay can be modeled as sigmoid functions of the number of VMs on a physical server 10
11. How close are the sigmoid functionsThe table shows the R-square values of different games/VMThe figure shows the curve fitting results with different number of VMs 11
12. Problem formulationObjective Function: Maximize ProfitsConstraint: QoE DegradationFrame Per SecondDelayDecision variable: 12
13. Other Constraints 13
14. outlineIntroductionProblem FormulationQDH AlgorithmTestbedTrace-Driven SimulationsMeasurement Study of Modern GPU14
15. Quality-driven heuristic (QDH)Consolidate more VMs on a serverDo not exceed the user-specified maximal tolerable QoE degradation Pseudocode15
16. QDH’ – Alternative algorithmAlternative Formulation and Algorithms for Closed SystemsObjective Function: Pseudocode16
17. outlineIntroductionProblem FormulationQDH AlgorithmTestbedTrace-Driven SimulationsMeasurement Study of Modern GPU17
18. Components of our systemBrokerVMware vCenter 5.1Single-Sign-On: authentication Inventory Service: managing/monitoring the VMs on ESXi serversPhysical ServersVMware ESXi 5.1GA Client/ServersGA is the first open source cloud gaming system Each VM host one GA server18
19. Flow of our systemGA client sends the account and password from gamer to brokerThe broker authenticates the gamerGA client sends the user-specified game to the brokerThe broker determines where to create a new VM and instructs the chosen physical server to launch a VMPhysical server sends the VM’s IP address to BrokerBroker Forwards the IP address to GA clientGA client connects to the GA serverGA server launches the gameGA server streams the game to gamer19
20. SET up – hardwarePhysical ServersCPU: i5 3.5 GHzGPU: Nvidia Quadro 6000Memory: 16GBBrokerCPU: i7 3.2 GHzMemory: 16GBClientsCPU: i5Memory: 4GBVMsEqually allocate the CPU and Memory to VMs20
21. Set up - scenarioJoin and leaving a game session with D% and (1-D)% probability (D% = 90%) in every minutesGame: Limbo, PSR, and NormandyRandomly select gameUp to 2 VMs for each physical serverTotal time: 15 minutes21
22. Practical concernMigration time of 20, 30, and 40 GB VM images are about 6, 9, 11 minutesDouble resources will be consumed between t1 to t3 while we do live migrationDecrease the profitsDecrease the QoEHence, we consider an migrationless version of proposed QDH/QDH’ algorithmsOnly place the VMs of incoming gamers to avoid the degradation caused by migration timeLive Migration22
23. Migrationless Algorithms are betterOutperforms QDH up to 396$ and 4% QoE 23
24. Performance of migrationless algorithms with different migration overhead25% migration overhead will achieve the same profitDue to the increasingly higher computing power, the migration overhead will be gradually reduced and the performance gains may be diminishing24
25. Conclusion of testbedAt this time, we do not consider QDH algorithm in the rest of the thesisQDH algorithm will be useful in the future while the migration time is reduced25
26. outlineIntroductionProblem FormulationQDH AlgorithmTestbedTrace-Driven SimulationsMeasurement Study of Modern GPU26
27. Set upNetwork latencies: KINGServer IP: OnLvie data centerClient IP: BitTorrentWoW tracesArrival time and leaving time of gamersGamesLimbo, PSR, and NormandyComputer:CPU: I7-3770 3.2 GHzMemory: 16GBVMs: Equally allocate the CPU and Memory to VMs27
28. Baseline algorithmLocation Based Placement (LBP) algorithm: LBP places each VM on a random game server that is not fully loaded and the data center geographically closest to the gamer28
29. Results of Provider centric algorithmEarn more money, up to 20+ thousand dollarsShutdown more servers(a) (b) Simulation results with WoW traces: (a) net profits and (b) used servers29
30. Results of gamer centric algorithmOutperform LBP algorithm up to 130% QoE30
31. Impact of number of gamersThe figure shows that more gamers lead to higher profits and lower QoE levels, and QDHL/QDH′ L successfully achieve their design objectives31
32. Running timeThe efficient algorithms terminate in < 2.5 s on a commodity PC even for large services with 20000 servers and 40000 gamers32
33. outlineIntroductionProblem FormulationQDH AlgorithmTestbedTrace-Driven SimulationsMeasurement Study of Modern GPU33
34. Modern GPUNvidia Quadro 6000Released in 2010Nvidia K2Released in 2013Support vGPUEach instance can be configured to: Pass-throughvGPU with up to 2,4,or 8 VMsSpecifications of Two GPUs34
35. Set upCloud gaming server: OS: XenServer 6.2CPU: Xeon 2.1 GHz Memory: 64 GBVM:By default, the XenServer allocates 1 CPU core and 2GB memory to Dom0, which is responsible for managing VMsThe remaining CPU cores and memory are equally divided among the VMs running Windows 735
36. Workload Game:Limbo: scroll-based puzzle gameFear2: first person shooter gameLEGO Batman: action gameBenchmark:Sanctuary: GPU benchmarkCadalyst: 2D versus 3DTinytask:A program to record the mouse and keyboard inputs of each gameWe record 3 minutes for each game and replay the same inputs to ensure fair comparisons36
37. Performance metricsFrame Per SecondThe number of rendered frames per secondContext SwitchThe number of context switches in Dom0CPU UtilizationThe CPU load of Dom0 (CPUdom0) and each VM (CPUvm).GPU UtilizationThe load of GPUs37
38. Measurement UtilitiesFraps: To measure the FPS of the foreground windowSar: To measure the number of context switchesXentop: To measure the CPU utilization of Dom0 and VMsNvidia-smi: To measure the GPU utilization under vGPUGPU-Z: To measure the GPU utilization of pass-through GPUs38
39. Performance of two modern GPUsThis table shows that K2 outperforms Quadro 6000 with up to 3.87 times of FPS increasesScalability: FPS of K2 does not drop too much even with 8 VMsHuge edge of vGPU (mediated pass-through) over vSGA (software-based virtualization)we no longer consider Quadro 6000 and vSGA in the rest of this thesisAchieved frame rates on two considered GPUs39
40. Independence of the two k2 GPU instancesPerformances of passThrough are the same in different configurations. vGPU has the same observationThe results show the two instances are independentSanctuary Scores in FPS from Diverse GPU Configurations40
41. Shared GPUs may outperform dedicated GPUsvGPU results in higher FPS than pass-through when executing Limbo and Fear2 Comparing the pass-through and vGPU: resulting FPS 41
42. 2D/3D performances vGPU2 outperforms pass-through …All 2D operations up to 15%Part of 3D operationsSimilar observations are also true for vGPU4 and vGPU8(a) (b) Comparing the pass-through and vGPU: (a) 2D benchmark scores and(b) 3D benchmark scores.42
43. Consolidation overheadLimbo does not suffer from consolidation overhead, while all other games/benchmark doGPU consolidation overhead: resulting FPS 43
44. Consolidation overhead caused by…The figure shows that Sanctuary is bounded by GPU, while Fear2 and Batman are bounded by CPUdom0Allocating more CPU cores to Dom0 to alleviate the high consolidation overhead for more complex games.(a) (b) GPU consolidation overhead: (a) fully loaded time ratio from vGPU8, and (b) CPUvm from vGPU8. 44
45. End-to-end cloud gaming performanceOnly 1 VMOpen source cloud gaming system: Gaming anywhereNot good-enough quality which between 20~42 fpsEnd-to-end performance of a cloud game platform: resulting FPS45
46. Reason of the low gaming qualityReal-time video encoding relies on computing power of CPULeverage the hardware codec on K2 GPU to improve it(a) (b) End-to-end performance of a cloud game platform: (a) CPUvm utilization with pass-through GPU and (b) CPUvm utilization with vGPU2.46
47. ConclusionVM placement algorithms [NetGames’13 short and IEEE TCC’14]:Migrationless algorithms outperform the state-of-the-art algorithm up to 20+ thousand dollars in net profits and 130% performance in QoEThe efficient algorithms terminate in 2.5 s on 20000 servers and 40000 clientsGPU measurement [MM’14 short under review]:Shared GPUs may outperform dedicated GPUsShared GPUs are rather scalable to the number of VMsModern GPUs can be shared by VMs running GPU-intensive computer games47
48. Future workHardware codecLeveraging the hardware H.264 codecs to improve the performance of real-time encoding More comprehensive system modelsOther types of resourcesHeterogeneous server types48
49. Q&A49