1. May 25, 2016 1 Our team: Undergrad Tim Hegeman, Stefan Hugtenburg, Jesse Donkevliet … Grad Siqi Shen, Guo Yong, Nezih Yigitbasi Staff Henk Sips, Dick Epema, Alexandru Iosup, Otto Visser Collaborators Ion Stoica and the Mesos team (UC Berkeley), Thomas Fahringer, Radu Prodan, Vlad Nae (U. Innsbruck), Nicolae Tapus, Mihaela Balint, Vlad Posea, Alexandru Olteanu (UPB), Derrick Kondo (INRIA), Assaf Schuster, Mark Silberstein, Orna Ben-Yehuda (Technion), Kefeng Deng (NUDT, CN),... Introduction to Cloud Computing Alexandru Iosup Parallel and Distributed Systems Group Delft University of Technology The Netherlands SPEC RG Cloud Meeting

2. May 25, 2016 2

3. 3

4. What is Cloud Computing? 3. A Useful IT Service “Use only when you want! Pay only for what you use!” May 25, 2016 4 Q: What do you use? Q: Why not this level?

5. May 25, 2016 5 Agenda 1.What is Cloud Computing? 2.IaaS Clouds, the Core Idea 3.The IaaS Owner Perspective 4.The IaaS User Perspective 5.Reality Check 6.Conclusion

6. IaaS Cloud Computing VENI – @larGe: Massivizing Online Games using Cloud Computing

7. Joe Has an Idea ($$$) (Source: A. Antoniou, MSc Defense, TU Delft, 2012. Original idea: A. Iosup, 2011.) MusicWave

8. Big up-front commitment Load variability: NOT supported Solution #1 Buy or Rent … 10% (Source: A. Antoniou, MSc Defense, TU Delft, 2012. Original idea: A. Iosup, 2011.)

9. Solution #2 Deploy on IaaS Cloud (Source: A. Antoniou, MSc Defense, TU Delft, 2012. Original idea: V. Nae, 2008.) Q: So are we just shifting the problem to somebody else, that is, the IaaS cloud owner? NO big up-front commitment Load variability: supported

10. Inside an IaaS Cloud Data Center (Source: A. Antoniou, MSc Defense, TU Delft, 2012. Original idea: A. Iosup, 2011.)

11. Time and Cost Sharing Among Users User C User B MusicWave (Source: A. Antoniou, MSc Defense, TU Delft, 2012.)

12. Main Characteristics of IaaS Clouds 1.On-Demand Pay-per-Use 2.Elasticity (cloud concept of Scalability) 3.Resource Pooling 4.Fully automated IT services 5.Quality of Service May 25, 2016 12

13. May 25, 2016 13 Agenda 1.What is Cloud Computing? 2.IaaS Clouds, the Core Idea 3.The IaaS Owner Perspective: How to Deploy a Cloud? 4.The IaaS User Perspective 5.Reality Check 6.Conclusion

14. IaaS Cloud Deployment Models Private On-premises Public Off-premises Hybrid (Source: A. Antoniou, MSc Defense, TU Delft, 2012. Original idea: Mell and Grance, NIST Spec.Pub. 800-145, Sep 2011.)

15. Resource Sharing Models MusicWave May 25, 2016 15 MusicWave OtherApp Space-SharingTime-Sharing IaaS Clouds MusicWave OtherApp Q: Which one is better? Grids Host OS OtherApp

16. Virtualization May 25, 2016 16 Virtualization Host OS MusicWaveOtherApp Q: What to do now? Guest OS Virtual Resources VM Instance Applications Guest OS Virtual Resources VM Instance Applications Q: What is the problem?

17. May 25, 2016 Virtualization and The Full IaaS Stack 17 Guest OS Virtual Resources VM Instance Applications Physical Infrastructure Virtual Infrastructure Manager Virtual Machine Manager Guest OS Virtual Resources VM Instance Applications Virtual Machine Manager Guest OS Virtual Resources VM Instance Applications

18. The Virtual Machine Lifecycle May 25, 2016 18 (Source: A. Antoniou, MSc Defense, TU Delft, 2012.) Q: Is this fair?

19. Use Case: Amazon Elastic Compute Cloud (EC2) Prominent IaaS provider Datacenters all over the world Many VM instance types Per-hour charging May 25, 2016 19 InstanceCapacityUS$/hour m1.small0.10 m1.large0.38 c1.xlarge0.76

20. May 25, 2016 20 Agenda 1.What is Cloud Computing? 2.IaaS Clouds, the Core Idea 3.The IaaS Owner Perspective 4.The IaaS User Perspective: How to Use Clouds? How to Choose Clouds? 5.Reality Check 6.Conclusion

21. Workload May 25, 2016 21 MusicWaveOtherApp Time MusicWave OtherApp Load = 4 RunTime= 6

22. Use Case: Workloads of Zynga (Massively Social Gaming) May 25, 2016 22 Sources: CNN, Zynga. Source: InsideSocialGames.com “Zynga made more than $600M in 2010 from selling in-game virtual goods.” S. Greengard, CACM, Apr 2011 Selling in-game virtual goods: “Zynga made est. $270M in 2009 from.” http://techcrunch.com/2010/ 05/03/zynga-revenue/ http://techcrunch.com/2010/ 05/03/zynga-revenue/

23. Use Case: Workloads of Zynga (Massively Social Gaming) Load can grow very quickly May 25, 2016 23 Load

24. Provisioning and Allocation of Resources May 25, 2016 24 Load Time ProvisioningAllocation

25. Provisioning and Allocation of Resources May 25, 2016 25 Load Time ProvisioningAllocation Q: What is the interplay between provisioning and allocation?

26. Provisioning and Allocation Policies May 25, 2016 26 Where? When? How many? Time Load ProvisioningAllocation From where? Which type? etc. When? etc. Q: How many policies exist?Q: How to select a policy? (Source: A. Antoniou, MSc Defense, TU Delft, 2012.)

27. Use Case: Two Provisioning Policies, Compared May 25, 2016 27 Startup OnDemand Villegas, Antoniou, Sadjadi, Iosup. An Analysis of Provisioning and Allocation Policies for Infrastructure- as-a-Service Clouds, (submitted). PDS Tech.Rep.2011-009

28. Use Case: Two Provisioning Policies, Compared Metrics for comparison Job Slowdown (JSD ): Ratio of actual runtime in the cloud and the runtime in a dedicated non-virtualized environment Charged Cost (C c ) Utility (U ) May 25, 2016 28 Q: Charged cost vs Total RunTime? Villegas, Antoniou, Sadjadi, Iosup. An Analysis of Provisioning and Allocation Policies for Infrastructure- as-a-Service Clouds, (submitted). PDS Tech.Rep.2011-009

29. Use Case: Two Provisioning Policies, Compared Workloads May 25, 2016 29 Uniform IncreasingBursty Villegas, Antoniou, Sadjadi, Iosup. An Analysis of Provisioning and Allocation Policies for Infrastructure- as-a-Service Clouds, (submitted). PDS Tech.Rep.2011-009

30. SystemHardwareVIMHypervisorMax VMs DAS4/Delft20 Dual quad- core 2.4 GHz 24 GB RAM 2x1 TB storage 64 FIU7 Pentium 4 3.0 GHz 5 GB RAM 340 GB Storage 7 Amazon EC2unkown/various-20 Use Case: Two Provisioning Policies, Compared Environments May 25, 2016 30 Villegas, Antoniou, Sadjadi, Iosup. An Analysis of Provisioning and Allocation Policies for Infrastructure- as-a-Service Clouds, (submitted). PDS Tech.Rep.2011-009

31. Use Case: Many Provisioning Policies, Compared Job Slowdown (JSD) May 25, 2016 31 Q: Why is OnDemand worse than Startup? A: waiting for machines to boot

32. Use Case: Many Provisioning Policies, Compared Charged Cost (C c ) May 25, 2016 32 Q: Why is OnDemand worse than Startup? A: VM thrashing Q: Why no OnDemand on Amazon EC2?

33. Use Case: Many Provisioning Policies, Compared Utility (U ) 33

34. May 25, 2016 34 Agenda 1.What is Cloud Computing? 2.IaaS Clouds, the Core Idea 3.The IaaS Owner Perspective 4.The IaaS User Perspective 5.Reality Check: Who Uses Public Commercial Clouds? 6.Conclusion

35. May 25, 2016 35 The Real IaaS Cloud “The path to abundance” On-demand capacity Cheap for short-term tasks Great for web apps (EIP, web crawl, DB ops, I/O) “The killer cyclone” Not so great performance for scientific applications (compute- or data-intensive) http://www.flickr.com/photos/dimitrisotiropoulos/4204766418/ Tropical Cyclone Nargis (NASA, ISSS, 04/29/08) VS May 25, 2016

36. 36 (Source: http://www.cca08.org/files/slides/w_vogel.pdf)

37. Zynga zCloud: Hybrid Self-Hosted/EC2 After Zynga had large scale More efficient self-hosted servers Run at high utilization Use EC2 for unexpected demand May 25, 2016 37 (Sources: http://seekingalpha.com/article/609141-how-amazon-s-aws-can-attract-ugly-economics and http://www.undertheradarblog.com/blog/3-reasons-zynga-is-moving-to-a-private-cloud/)http://seekingalpha.com/article/609141-how-amazon-s-aws-can-attract-ugly-economics

38. Other Cloud Customers 218 virtual CPUs 9TB/2TB block/S3 storage 6.5TB/2TB I/O per month May 25, 2016 38 (Source: http://markbuhagiar.com/technical/businessinthecloud/)

39. May 25, 2016 39 Agenda 1.What is Cloud Computing? 2.IaaS Clouds, the Core Idea 3.The IaaS Owner Perspective 4.The IaaS User Perspective 5.Reality Check 6.Conclusion

40. May 25, 2016 40 Conclusion Take-Home Message Cloud Computing = IaaS + PaaS + SaaS Core idea = lease vs self-own On-Demand, Pay-per-Use, Elastic, Pooled, Automated, QoS The Owner Perspective Time-Sharing Virtualization The User Perspective Variable workloads Provisioning and Allocation policies Reality Check: 100s of users http://www.flickr.com/photos/dimitrisotiropoulos/4204766418/

41. May 25, 2016 41 Thank you for your attention! Questions? Suggestions? Observations? Alexandru Iosup A.Iosup@tudelft.nl http://www.pds.ewi.tudelft.nl/~iosup/ (or google “iosup”) Parallel and Distributed Systems Group Delft University of Technology A.Iosup@tudelft.nl http://www.pds.ewi.tudelft.nl/~iosup/ - http://www.st.ewi.tudelft.nl/~iosup/research.htmlhttp://www.st.ewi.tudelft.nl/~iosup/research.html - http://www.st.ewi.tudelft.nl/~iosup/research_cloud.htmlhttp://www.st.ewi.tudelft.nl/~iosup/research_cloud.html - http://www.pds.ewi.tudelft.nl/http://www.pds.ewi.tudelft.nl/ More Info: Do not hesitate to contact me…

42. … And Now For Something Different Ongoing projects in Cloud Computing at TUD May 25, 2016 42

43. Massivizing Social Games: Distributed Computing Challenges and High Quality Time – A. Iosup 43 Cloudification: PaaS for MSGs (Platform Challenge) Build Social Gaming platform that uses (mostly) cloud resources Close to players No upfront costs, no maintenance Compute platforms: multi-cores, GPUs, clusters, all-in-one! Performance guarantees Hybrid deployment model Code for various compute platforms—platform profiling Load prediction miscalculation costs real money What are the services? Vendor lock-in? My data

44. Data, Data, Data Understanding the Behavior of Players in Online Social Games May 25, 2016 44 460 73 66 596 871 198 1075 1004 32 127 51 240 (w/ E.Dias, A. Olteanu, F. Kuipers) Iosup et al. An Analysis of Implicit Social Networks in Multiplayer Online Games.IEEE Internet Computing

45. 45 Using data centers for dynamic resource allocation Main advantages: 1. Significantly lower over-provisioning 2. Efficient coverage of the world is possible Proposed hosting model: dynamic Massive join Massive leave Nae, Iosup, Prodan. Dynamic Resource Provisioning in Massively Multiplayer Online Games. IEEE TPDS 2011

46. 46 Resource Provisioning and Allocation Static vs. Dynamic Provisioning 250% 25% [Source: Nae, Iosup, and Prodan, ACM SC 2008] Over-provisioning = WASTE (%) Nae, Iosup, Prodan. Dynamic Resource Provisioning in Massively Multiplayer Online Games. IEEE TPDS 2011

47. Old scheduling aspects Workloads evolve over time and exhibit periods of distinct characteristics No one-size-fits-all policy: hundreds exist, each good for specific conditions Data centers increasingly popular (also not new) Constant deployment since mid-1990s Users moving their computation to IaaS-cloud data centers Consolidation efforts in mid- and large-scale companies New scheduling aspects New workloads New data center architectures New cost models Developing a scheduling policy is risky and ephemeral Selecting a scheduling policy for your data center is difficult Combining the strengths of multiple scheduling policies is … Why Portfolio Scheduling?

48. What is Portfolio Scheduling? In a Nutshell, for Data Centers Create a set of scheduling policies Resource provisioning and allocation policies, in this work Online selection of the active policy, at important moments Periodic selection, in this work Same principle for other changes: pricing model, system, … Deng, Song, Ren, Iosup: Exploring portfolio scheduling for long-term execution of scientific workloads in IaaS clouds. SC 2013: 55

49. Performance Evaluation 1) Effect of Portfolio Scheduling (1) Portfolio scheduling is 8%, 11%, 45%, and 30% better than the best constituent policy A portfolio scheduler can be better than any of its constituent policies Q: What can prevent a portfolio scheduler from being better than any of its constituent policies? Deng, Song, Ren, Iosup: Exploring portfolio scheduling for long-term execution of scientific workloads in IaaS clouds. SC 2013: 55

50. Performance Evaluation 1) Effect of Portfolio Scheduling (2) Job selection policies such as UNICEF and LXF that favor short jobs have the best performance Q: How well do you think a single (provisioning, job selection, VM selection) policy would perform? Will it be dominant? (Rhetorical)

51. Massivizing Social Games: Distributed Computing Challenges and High Quality Time – A. Iosup 51 Mobile Social Gaming and the SuperServer (Platform Challenge) Support Social Gaming on mobiles Mobiles everywhere (2bn+ users) Gaming industry for mobiles is new Growing Market SuperServer to generate content for low-capability devices? Battery for 3D/Networked games? Where is my server? (Ad-hoc mobile gaming networks?) Security, cheat-prevention

52. Extending the Capabilities of Mobile Devices through Cloud Offloading … with Application to Online Social Games Metrics: processing time packet size inter-arrival rate Design & Implementation: based on OpenTTD repeatability offloading mechanisms instrumentation for metrics

53. 53 Social Everything! So Analytics Social Network=undirected graph, relationship=edge Community=sub-graph, density of edges between its nodes higher than density of edges outside sub-graph (Analytics Challenge) Improve gaming experience Ranking / Rating Matchmaking / Recommendations Play Style/Tutoring Self-Organizing Gaming Communities Player Behavior

54. Asterix B-tree How to Analyze? Ecosystems of Big-Data Programming Models Dremel Service Tree SQLHivePigJAQL MapReduce ModelAlgebrixPACT MPI/ Erlang LFSLFS NepheleHyracksDryadHadoop/ YARN Haloop DryadLINQScope Pregel HDFS AQL CosmosFS Azure Engine Tera Data Engine Adapted from: Dagstuhl Seminar on Information Management in the Cloud, http://www.dagstuhl.de/program/calendar/partlist/?semnr=11321&SUOG http://www.dagstuhl.de/program/calendar/partlist/?semnr=11321&SUOG Azure Data Store Tera Data Store Storage Engine Execution Engine Voldemort High-Level Language Programming Model GFS BigQueryFlume Flume Engine S3 Dataflow Giraph SawzallMeteor * Plus Zookeeper, CDN, etc.

55. May 25, 2016 55 Benchmarking suite Platforms and Process Platforms Process Evaluate baseline (out of the box) and tuned performance Evaluate performance on fixed-size system Future: evaluate performance on elastic-size system Evaluate scalability YARN Giraph Guo, Biczak, Varbanescu, Iosup, Martella, Willke. How Well do Graph-Processing Platforms Perform? An Empirical Performance Evaluation and Analysis

56. May 25, 2016 56 BFS: results for all platforms, all data sets No platform can runs fastest of every graph Not all platforms can process all graphs Hadoop is the worst performer Guo, Biczak, Varbanescu, Iosup, Martella, Willke. How Well do Graph-Processing Platforms Perform? An Empirical Performance Evaluation and Analysis

57. May 25, 2016 57 Giraph: results for all algorithms, all data sets Storing the whole graph in memory helps Giraph perform well Giraph may crash when graphs or messages become larger Guo, Biczak, Varbanescu, Iosup, Martella, Willke. How Well do Graph-Processing Platforms Perform? An Empirical Performance Evaluation and Analysis

58. May 25, 2016 58 Conclusion and ongoing work Performance is f(Data set, Algorithm, Platform, Deployment) Cannot tell yet which of (Data set, Algorithm, Platform) the most important (also depends on Platform) Platforms have their own drawbacks Some platforms can scale up reasonably with cluster size (horizontally) or number of cores (vertically) Ongoing work Benchmarking suite Build a performance boundary model Explore performance variability Guo, Biczak, Varbanescu, Iosup, Martella, Willke. How Well do Graph-Processing Platforms Perform? An Empirical Performance Evaluation and Analysis.IPDPS http://bit.ly/10hYdIU

59. DotA communities Players are loosely organised in communities Operate game servers Maintain lists of tournaments and results Publish statistics and rankings on websites Dota-League: players join a queue and matchmaking forms teams DotAlicious: players can choose which match/team to join R. van de Bovenkamp, S. Shen, A. Iosup, F. A. Kuipers: Understanding and recommending play relationships in online social gaming. COMSNETS 2013: 1-10

60. Relationships in the gaming graph Players who regularly play together in DotAlicious do so in more diverse combinations than in Dota-League Contrary to Dota-League, DotAlicious players tend to play on the same side: playing together intensifies the social bond Winning together increases friendship relationships, while loosing together weakens friendship relationships Small clusters of friends with very strong social ties exist R. van de Bovenkamp, S. Shen, A. Iosup, F. A. Kuipers: Understanding and recommending play relationships in online social gaming. COMSNETS 2013: 1-10

61. Matchmaking application Replay match list, but also consider clusters in gaming graph Scoring methodology: Points per cluster: Number of players in the match that are part of the same cluster Excluding largest cluster of the network and clusters of size 1 PlayerCluster a1 b2 c1 d3 e4 PlayerCluster f2 g5 h3 i6 j3 Team 1Team 2 ClusterPoints 12 22 33 40 50 Total of 7 points for this match R. van de Bovenkamp, S. Shen, A. Iosup, F. A. Kuipers: Understanding and recommending play relationships in online social gaming. COMSNETS 2013: 1-10

62. Results matchmaking Can already improve original matchmaking algorithm for all gaming graphs! Dota-LeagueDotAlicious R. van de Bovenkamp, S. Shen, A. Iosup, F. A. Kuipers: Understanding and recommending play relationships in online social gaming. COMSNETS 2013: 1-10 Can do much better than random matchmaking