1. Shanjiang Tang, Bu-Sung Lee, Bingsheng He, Haikun Liu School of Computer Engineering Nanyang Technological University Long-Term Resource Fairness Towards Economic Fairness on Pay-as-you-use Computing Systems

2. Pay-As-You-Use is Pervasive Charge users based on the amount of resources used over time (e.g., Hourly). Advantages – Elasticity – Flexibility – Cost efficiency Pay-as-you-use is becoming common and popular. – Supercomputing, Cloud Computing 2

3. Twitter’s Cluster One week data from Twitter production cluster [Delimitrou et. Al. ASPLOS’14] Resource Utilization = User resource demands are heterogeneous. – Users have different demands. – A user’s demand is changing over time.  Static provisioning/partitioning causes underutilization. Resource utilization is a critical problem in such pay-as-you-use environments. – Providers waste resources (  waste investment and lose profit). – Users waste money. 3

4. Resource Sharing can improve resource utilization. – Allow underloaded users to release resources to other users. – Allow overloaded users to temporarily use more resources (from others).  Reduce the idle resources at runtime.  Resolve resource contention across users. What about fairness? – If the fairness is not solved, resource sharing is unlikely to achieve in pay-as-you-use environments. To Share or Not To Share? 4

5. Pay-as-you-use Fairness: Resource-as-you-pay The total resources a user gained should be proportional to her payment. This is a Service-Level Agreement (SLA). 60 $ 40 $ A: B: 60% 40% Resource Service A A B B Resource Service = Resources-per-time X service time 5

6. Example: Amazon EC2 Reserved Instance – Pay one-time fee for a long time in advance – Get discount on the hourly charge over on-demand one. E.g., m3.xlarge: on-demand($0.28/h), reserved instance (0.06/h) Annual UtilizationMedium Utilization RIHeavy Utilization RI 20%-8%-99% 40%30%1% 60%43%34% 80%50% 100%54%60% Table: 3-year RI Percentage Saving Over on-Demand Comparison [data from aws] 6

7. Desirable Properties in Pay-as- you-use Computing Resource-as-you-pay Fairness Guarantee Non-Trivial Workload Incentive – User should submit non-trivial workload – Be willing to yield to others when no need. – Improve cost-efficiency. Truthfulness – Users cannot get benefits by cheating. – Be honest and friendly to each other. 7

8. Fair Policy in Existing Systems State-of-the-art: Max-min fairness – Select the user with the minimum allocation/share ratio every time. – Consider the present requirement only (memoryless). Memoryless fairness has severe problems in pay-as-you-use environments, violating the following properties: – Resource-as-you-pay fairness guarantee. – Non-Trivial workload incentive and sharing incentive. – Truthfulness (Users may get benefits by cheating). 8

9. Problems with MemoryLess Fairness Resource-as-you-pay Fairness Problem – E.g., A, B equally pay for total resource of 100 units. Time New Demand AB t1 20100 A A B B Accumulate Resource Usage: Accumulate Resource Usage: 20 80 20 80 Unsatisfied Demand A A B B 0 0 20 9 Current Allocation at t1:

10. Problems with MemoryLess Fairness Resource-as-you-pay Fairness Problem – E.g., A, B equally pay for total resource of 100 units. Time New Demand AB t1 20100 t2 4060 A A B B Accumulate Resource Usage: Accumulate Resource Usage: 40 60 140 A A B B Unsatisfied Demand 0 0 20 10 Current Allocation at t2:

11. Problems with MemoryLess Fairness Resource-as-you-pay Fairness Problem – E.g., A, B equally pay for total resource of 100 units. Time New Demand AB t1 20100 t2 4060 t3 8050 A A B B Accumulated resource usage: 50 110 190 A A B B Unsatisfied Demand 30 20 11 Current Allocation at t3:

12. Problems with MemoryLess Fairness Resource-as-you-pay Fairness Problem – E.g., A, B equally pay for total resource of 100 units. Time New Demand AB t1 20100 t2 4060 t3 8050 t4 6050 A A B B Accumulated resource usage: 50 160 240 A A B B Unsatisfied Demand 40 20 Existing Fair Policy fails to satisfy Resource-as-you-pay fairness!!! 12 Current Allocation at t4:

13. MemoryLess Fairness Violates Sharing Incentives Non-trivial workload and sharing incentive Problem – Yielding resources to others have no benefits. – Suppose A, B, and C equally pay for total resource of 100 units. A has 13 idle resource units. In that case, A can be selfish, either idle or running trivial workloads. CPU 20 33 A:A: A:A: B:B: B:B: C:C: C:C: A’s idle resource 13

14. Cheating User Benefits on MemoryLess Fairness Truthfulness Problem – Suppose A, B, C equally pay for a cluster of 100 units, with true demand to be 33, 21 and 80, respectively. – Case 1: all are honest. – Case 2: User A cheats and claims the demand to be 40. 14 33 21 33 A:A: A:A: B:B: B:B: C:C: C:C: 6 6 6 6 A’s cheating get benefits 33 21 33 A:A: A:A: B:B: B:B: C:C: C:C: 12 Case 1: A is honest Case 2: A is cheating

15. Our Work Challenges: can we find a fair sharing policy that satisfies the following properties? – Resource-as-you-pay fairness – Non-trivial workload and sharing incentives – Truthfulness Our Solution: Long-Term Resource Fairness – Ensure resource fairness over a period of time. – With historical information considered. 15

16. Long-Term Resource Fairness Basic Concept: Loan agreement (Lending w/o interests) – When resources are not needed, users can lend the resources to others. – When more resources are needed, others should give back.  Benefit others and user herself. 16

17. Long-Term Resource Fairness Satisfy Pay-as-you-use Fairness Time New Demand AB t1 20100 A A B B Accumulated resource usage: 20 80 20 80 Unsatisfied Demand A A B B 0 0 20 17 Current Allocation at t1: A A B B Lend Resources: 30 -30

18. Long-Term Resource Fairness Satisfy Pay-as-you-use Fairness Time New Demand AB t1 20100 t2 4060 A A B B Accumulated resource usage: 40 60 140 A A B B Unsatisfied Demand 0 0 20 18 Current Allocation at t2: A A B B Lend Resources: 40 -40

19. Long-Term Resource Fairness Satisfy Pay-as-you-use Fairness Time New Demand AB t1 20100 t2 4060 A A B B Accumulated resource usage: 40 60 140 A A B B Unsatisfied Demand 0 0 20 19 Current Allocation at t2: A A B B Lend Resources: 40 -40 t3 8050

20. Long-Term Resource Fairness Satisfy Pay-as-you-use Fairness Time New Demand AB t1 20100 t2 4060 t3 8050 A A B B Accumulated resource usage: 80 20 140 160 A A B B Unsatisfied Demand 0 0 50 20 Current Allocation at t3: A A B B Lend Resources: 10 -10

21. Long-Term Resource Fairness Satisfy Pay-as-you-use Fairness Time New Demand AB t1 20100 t2 4060 t3 8050 A A B B Accumulated resource usage: 80 20 140 160 A A B B Unsatisfied Demand 0 0 50 21 Current Allocation at t3: A A B B Lend Resources: 10 -10 t4 6050

22. Long-Term Resource Fairness Satisfy Pay-as-you-use Fairness Time New Demand AB t1 20100 t2 4060 t3 8050 t4 6050 A A B B Accumulated resource usage: 60 40 200 A A B B Unsatisfied Demand 0 0 60 Long-Term Resource Fairness satisfy Resource-as-you-pay fairness. 22 Current Allocation at t4: A A B B Lend Resources: 0 0 0 0

23. Other Properties of Long-Term Resource Fairness Satisfy non-trivial workload and sharing incentives – Running trivial workload can waste money. – Not sharing idle resource can waste money. Users cannot get benefits by lying (strategy proof). 23 Proof sketches are in the paper.

24. LTYARN Implement Long-Term Resource Fairness in YARN – Extend memoryless max-min fairness to long-term max- min fairness. – Add a few components into resource manager Support full long-term and time window-based requirements. Currently support a single resource type (main memory). 24

25. LTYARN Design Quantum Updater (QU) – Estimates task execution time. – Updates the resource usage history periodically. Resource Controller (RC) – Manages and updates resource for each queue. Resource Allocator (RA) – Performs long-term resource allocation. – Runs when there are pending tasks and idle resources. 25

26. Evaluation A Hadoop Cluster – 10 nodes, each with two Intel X5675 CPUs (6 cores per CPU with 3.07 GHz), 24GB DDR3 memory, 56GB hard disks. – YARN-2.2.0, configured with 24GB memory per node. Macro-benchmarks – Synthetic Facebook Workload – Purdue Workload – HIVE/TPC-H – Spark 26 Detailed setups are in the paper.

27. Metrics Evaluation metrics – Fairness degree for each user (>1 for sharing benefits; <1 for sharing loss) – Resource-as-you-pay fairness – Application performance Benchmark scenario – The four macro benchmarks equally share the cluster. – Each benchmark runs in a separate queue. – Window size =1 day. 27

28. Sharing Benefit/Loss LTYARN enables sharing benefits. Sharing benefit degree: the degree of benefits under sharing cluster over non-sharing case Sharing loss degree : the degree of losses under sharing cluster over non-sharing case (a). YARN 28 (b). LTYARN

29. Sharing Benefit/Loss LTYARN enables sharing benefits for all applications. (b). LTYARN 29 (a). YARN

30. Resource-as-you-pay Fairness Results LTYARN achieves resource-as-you-pay fairness. 30

31. Performance Results 31 Sharing always achieves a better performance. Long-term fairness is comparable to memory-less fairness (max-min).

32. Conclusions Max-min resource fairness is memoryless and unsuitable for pay-as-you-use computing. We define long-term resource fairness that can satisfy the desirable properties. We develop LTYARN by integrating long-term resource fairness into YARN – Homepage: http://sourceforge.net/projects/ltyarn/http://sourceforge.net/projects/ltyarn/ 32

33. Future Work Implement Long-Term Resource Fairness in other systems/schedulers. – Mesos, Quincy, Choosy, etc. Extend Long-Term Resource Fairness for multi- resources: – CPU, memory, Network I/O, etc. 33

34. We are Hosting IEEE CloudCom 2014 in Singapore Deadline for paper submissions: July 31, 2014 Notification of Paper acceptance: September 2, 2014 Conference: December 15-18, 2014 34

35. Thanks! Question? 35