Predictive Elastic Database Systems. Rebecca Taft HPTS 2017

Transcription

1 Predictive Elastic Database Systems Rebecca Taft HPTS

2 Modern OLTP Applications Large Scale Cloud-Based Performance is Critical 2

3 Challenges to transaction performance: skew and workload variation Database Elasticity E-Store manage skew and react to variation P-Store predictive modeling for time variation 3

4 E-Store: Elastic Scaling to Adapt to Workload Uniform Workload, Increasing Load TransacHons Transaction Arrival Rate Part 1 Part 2 Part 3 4

5 E-Store: Elastic Scaling to Adapt to Workload Uniform Workload, Increasing Load 10 Transaction Arrival Rate TransacHons Part 1 Part 2 Part 3 4

6 E-Store: Elastic Scaling to Adapt to Workload Uniform Workload, Increasing Load 10 Transaction Arrival Rate TransacHons Part 1 Part 2 Part 3 Part 4 Part 5 4

7 E-Store: Elastic Scaling to Adapt to Workload Uniform Workload, Increasing Load 10 Transaction Arrival Rate TransacHons Part 1 Part 2 Part 3 Part 4 Part 5 4

8 E-Store: Elastic Scaling to Adapt to Workload Uniform Workload, Increasing Load TransacHons Transaction Arrival Rate Part 1 Part 2 Part 3 Part 4 Part 5 4

9 E-Store: Elastic Scaling to Adapt to Workload Uniform Workload, Increasing Load Skewed Workload 10 Transaction Arrival Rate 10 Transaction Arrival Rate TransacHons Part 1 Part 2 Part 3 Part 4 Part 5 TransacHons Part 1 Part 2 Part 3 4

10 E-Store: Elastic Scaling to Adapt to Workload Uniform Workload, Increasing Load Skewed Workload 10 Transaction Arrival Rate 10 Transaction Arrival Rate TransacHons Part 1 Part 2 Part 3 Part 4 Part 5 TransacHons Part 1 Part 2 Part 3 4

11 E-Store: Elastic Scaling to Adapt to Workload Uniform Workload, Increasing Load Skewed Workload 10 Transaction Arrival Rate 10 Transaction Arrival Rate TransacHons Part 1 Part 2 Part 3 Part 4 Part 5 TransacHons Part 1 Part 2 Part 3 4

12 E-Store Elastic DBMS Architecture Elastic Controller Load Monitor Planner Live Migration Shared Nothing DBMS (e.g. H-Store) 5

13 A Case Study: B2W Digital 6

14 3-Day Online Retail Database Workload 7

15 3-Day Online Retail Database Workload 7

16 3-Day Online Retail Database Workload 7

17 3-Day Online Retail Database Workload 7

18 3-Day Online Retail Database Workload 7

19 Elastic Scaling Adapts to Workload 8

20 Reactive Scaling Causes Latency Spikes 8

21 P-Store A Predictive Elastic DBMS 9

22 P-Store Architecture Predictive Controller Load Monitor Predictor Planner Live Migration Shared Nothing DBMS (e.g. H-Store) 10

23 Ideal Capacity Actual Servers Allocated Machine Capacity Demand 11

24 Machine Capacity Demand 12

25 Machine Capacity Demand? 13

26 Machine Capacity Demand Options 1. Add machine(s) 13

27 Machine Capacity Demand Options 1. Add machine(s) 13

28 Machine Capacity Demand Options 1. Add machine(s) 13

29 Machine Capacity Demand Options 1. Add machine(s) 2. Remove machine(s) 13

30 Machine Capacity Demand Options 1. Add machine(s) 2. Remove machine(s) 3. No change 13

31 Machine Capacity Demand Options 1. Add machine(s) 2. Remove machine(s) 3. No change 13

32 Machine Capacity Demand Effective Capacity Options 1. Add machine(s) 2. Remove machine(s) 3. No change 13

33 Machine Capacity Demand Effective Capacity Options 1. Add machine(s) 2. Remove machine(s) 3. No change Challenges 1. Time to Reconfigure 2. Effective Capacity 3. Cost 13

34 Time to Reconfigure Rate control: small chunks of data, spaced apart some time D minimum time to move entire database w/ no performance impact % of Data % of Data Nodes 14 Time = ¼ * D Nodes

35 Effective Capacity % of TransacHons What transaction rate can the system support? Max rate per machine: Q = % of TransacHons Nodes Nodes 15

36 Cost Nodes Allocated Time Cost = 7

37 Machine Capacity Demand Effective Capacity 4 machines at t = 9 Goal: find path with minimal cost such that eff. capacity > demand Cost of {path to A machines ending at time t} = Cost of {path to B machines ending at time t Time B A } + Cost B A 17

38 Dynamic Programming Algorithm for Planning Reconfigurations Complexity # time steps and max number of machines needed In practice: runtime < 1 second Greedy alternatives don t work 18

39 Machine Capacity Demand Effective Capacity 19

40 Machine Capacity Demand Effective Capacity Optimal Solution Total Cost: 30 19

41 Time Series Prediction Diurnal, periodic workload, some variation day to day Prediction must complete in seconds-to-minutes We use Sparse Periodic Auto-Regression (SPAR) Chen et al. NSDI 08 20

42 Evaluation Can we reduce resource usage? Can we prevent latency spikes? 21

43 Experiments 3 Days of B2W workload run on H-Store Cluster of 10 servers, 6 partitions per server About 1 GB of shopping cart and checkout data Track machines allocated, throughput, latency, reconfiguration time periods 22

44 Results Static, Peak Provisioning Machines Used: 10 23

45 Results Static, Average Provisioning Machines Used: 4 24

46 Results Reactive Scaling Avg. Machines Used:

47 Results P-Store with SPAR Avg. Machines Used:

48 Results P-Store with Exact Provisioning Avg. Machines Used:

49 Results Comparison: CDF of Top 1% of Latency 28

50 Summary: P-Store Evaluation Can we reduce resource usage? Saves 50% of computing resources compared to static allocation Can we prevent latency spikes? Superior performance compared to reactive approach On a real workload, P-Store reduces resource usage while keeping latency within application requirements 29

51 Summary Real database workloads are skewed and vary over time Elasticity enables management of skew and adaptation to load changes Predictive scaling improves performance during load changes compared to reactive scaling Rebecca Taft 30