Collaborative Query Coordination in Community-Driven Data Grids

Transcription

1 HPDC '09 Collaborative Query Coordination in Community-Driven Data Grids Tobias Scholl, Angelika Reiser, and Alfons Kemper Department of Computer Science, Technische Universität München Germany

2 Community-Driven Data Grids (HiSbase)

3 The AstroGrid-D Project German Astronomy Community Grid Funded by the German Ministry of Education and Research Part of D-Grid HPDC 2009 Collaborative Query Processing 3

4 Up-Coming Data-Intensive Applications Alex Szalay, Jim Gray (Nature, 2006): Science in an exponential world Data rates LHC Terabytes a day/night Petabytes a year LSST LOFAR Pan-STARRS LHC LOFAR HPDC 2009 Collaborative Query Processing 4

5 The Multiwavelength Milky Way HPDC 2009 Collaborative Query Processing 5

6 Research Challenges Directly deal with Terabyte/Petabyte-scale data sets Integrate with existing community infrastructures High throughput for growing user communities HPDC 2009 Collaborative Query Processing 6

7 Current Sharing in Data Grids Data autonomy Policies allow partners to access data Each institution ensures Availability (replication) Scalability Various organizational structures [Venugopal et al. 2006]: Centralized Hierarchical Federated Hybrid HPDC 2009 Collaborative Query Processing 7

8 Community-Driven Data Grids (HiSbase) HPDC 2009 Collaborative Query Processing 8

9 Community-Driven Data Grids (HiSbase) HPDC 2009 Collaborative Query Processing 9

10 Distribute by Region not by Archive! HPDC 2009 Collaborative Query Processing 10

11 Distribute by Region not by Archive! HPDC 2009 Collaborative Query Processing 11

12 Distribute by Region not by Archive! HPDC 2009 Collaborative Query Processing 12

13 Distribute by Region not by Archive! HPDC 2009 Collaborative Query Processing 13

14 Mapping Data to Nodes HPDC 2009 Collaborative Query Processing 14

15 Submission Characteristics Portal-based submission Browser in every researcher s "tool box Scalability depends on portal Institution-based submission All data nodes accept queries Submission via local data node HPDC 2009 Collaborative Query Processing 15

16 Coordinator Selection Strategies The node submitting the query SelfStrategy (SS) A node containing relevant data (region-based strategies) FirstRegionStrategy (FRS) SelfOrFirstRegionStrategy (SOFRS) CenterOfGravityStrategy (COGS) RandomRegionStrategy (RRS) HPDC 2009 Collaborative Query Processing 16

17 SelfStrategy (SS) HPDC 2009 Collaborative Query Processing 17

18 FirstRegionStrategy (FRS) HPDC 2009 Collaborative Query Processing 18

19 SelfOrFirstRegionStrategy (SOFRS) Combination from SelfStrategy and FirstRegionStrategy Submit node is coordinator if it covers data Avoids unnecessary data transport With many partitions and many nodes basically the same as FirstRegionStrategy (as probability of Self-case decreases) HPDC 2009 Collaborative Query Processing 19

20 CenterOfGravityStrategy (COGS) Further reduce amount of data shipping "Perfect spot for minimizing data transfer HPDC 2009 Collaborative Query Processing 20

21 RandomRegionStrategy (RRS) Select random relevant region Tradeoff between balancing coordination load and reducing data shipping Probability(a) = 2/9 Probability(b) = 5/9 Probability(c) = 2/ HPDC 2009 Collaborative Query Processing 21

22 Evaluation Coordination Strategies: SS, FRS, SOFRS, COGS, RRS Submission Strategies: portal-based, institution-based Observational data sets Two workloads SDSS query log (Q obs ) Synthetic (Q scaled ) Network size P obs Network traffic measurements Number of routed messages Coordination load balancing Throughput Measurements HPDC 2009 Collaborative Query Processing 22

23 Query Workloads HPDC 2009 Collaborative Query Processing 23

24 Routed Messages per Query (Q obs ) HPDC 2009 Collaborative Query Processing 24

25 Routed Messages per Query (Q scaled ) HPDC 2009 Collaborative Query Processing 25

26 Portal-based Coordination Load HPDC 2009 Collaborative Query Processing 26

27 Institution-based Coordination Load HPDC 2009 Collaborative Query Processing 27

28 Throughput Q obs Q scaled Throughput dependent on query complexity No clear winner in terms of throughput HPDC 2009 Collaborative Query Processing 28

29 Workload-Aware Data Partitioning Query skew (hot spots) triggered by increased interest in particular subsets of the data Two well-known query load balancing techniques: Data partitioning Data replication Finding trade-offs between both (see EDBT 09 paper) HPDC 2009 Collaborative Query Processing 29

30 Load Balancing During Runtime Complement workload-aware partitioning with runtime loadbalancing Short-term peaks Master-slave approach Load monitoring Long-term trends Based on load monitoring Histogram evolution HPDC 2009 Collaborative Query Processing 30

31 Related Work On-line load balancing Hundreds of thousands to millions of nodes Reacting fast Treating objects individually HiSbase HPDC 2009 Collaborative Query Processing 31

32 Who Is the Query Coordinator? Many challenges and opportunities in e-science for distributed computing and database research High-throughput data management Correlation of distributed data sources Collaborative Query Coordination Region-based strategies reduce number of messages Load balancing independent of submission characteristic HPDC 2009 Collaborative Query Processing 32

33 Special Thanks To Ella Qiu, University of British Columbia DAAD Rise Internship Support during implementation Initial measurements HPDC 2009 Collaborative Query Processing 33

34 Get in Touch Database systems group, TU München Web site: The HiSbase project Thank You for Your Attention HPDC 2009 Collaborative Query Processing 34