1. Analysis of Database Workloads on Modern Processors Advisor: Prof. Shan Wang P.h.D student: Dawei Liu Key Laboratory of Data Engineering and Knowledge Engineering MOE School of Information Renmin University of China

2. Outlines 1. Background 2. Motivation 3. Our research work 4. Future works

3. Background LAMA Project Goal Advanced issues of Massively Parallel Processing (MPP) databases Architecture and design aspects; Next generation memory oriented DB My Focus rgle Scale Data nagement La Ma Joint research with HP Lab China

4. Outlines 1. Background 3. Our research 4. Future works 2. Motivation

5. Motivation Continued evolution of hardware Processor

6. Motivation(cont.) Memory Larger and Larger Flash Memory

7. Cont. Traditional research Dedicate to I/O optimization Fail to utilize processor resources efficiently

8. Cont. Modern processors (Itanium II) multi-level memory hierarchies; superscalar out-of-order execution; multi-threading; multi-cores; Create opportunity for database performance improve.

9. Cont. Object Accurately characterizing workload behavior on modern processor Find out the bottleneck; Benefit Identify a set of characteristics; performance optimization Detailed issues ?

10. My P.h. D Track (1) Accurately characterize the database workloads on modern processors; (2) Investigating the MMDB workloads on modern processor; (3) Developing a specialized benchmark for MMDB

11. (1) Processor Issue Previous research [ * ] Conlusion DBMSs achieve low IPC (instructions-per-cycle) Processors are inefficiently used Platform Intel Pentium II / Pentium Pro ---------------------------------------------------------------------------------------------------------------------------------------- * A. Ailamaki, D. J. DeWitt, M. D. Hill, D. A. Wood. DBMSs on a Modern Processor: Where Does Time Go? In Proc. VLDB, 1999.

12. Cont. We are interested in DBMS on today’s processors Itanium II AMD Opteron (tm) Where does 8 years go ?

13. (2) Main Memory DB Issue Previous research DB: Disk Resident Databases (DRDB) Workload: TPC-C Current problems DB: Main Memory Databases (MMDB) Workload: TPC-H (compute intensive) The “moved up” on the memory hierarchy ; Larger and larger on-chip and off-chip caches ; Steady increased RAM;

14. (3) MMDB-Oriented Benchmark Performance evaluation OO1-Benchmark OO7-Benchmark obsolete Industrial standards How to benchmark memory database ? TPC Benchmark C TPC Benchmark H OLAPOLTP We found they are not approprite to benchmark MMDB

15. Outlines 1. Background 2. Motivation 4. Future works 3. Our research

16. Methodology Analysis framework Experiment study

17. Pipeline of modern processors

18. Query Execution Time Breakdown T Q = T C + T M + T B + T R − T OV L [ * ] T C : Useful computation time; T M : Stall time because of memory stalls; T B : Branch misprediction overhead; T R : Resource-related stalls; T OVL : Stall time can be overlapped * A. Ailamaki, D. J. DeWitt, M. D. Hill, D. A. Wood. DBMSs on a Modern Processor: Where Does Time Go? In Proc. VLDB, 1999.

19. Execution time components on Itanium II platform

20. Experimental setup Platform-specific hardware Software Experimental methodology

21. The Hardware Platform HP Integrity rx2620-2 server Itanium II based server Cache

22. Cache characteristics

23. Software and Methodology Calibrator (CWI * ) cache access and miss latency; main memory access latency; number of TLB levels ; each level’s TLB miss latency * Centrum voor Wiskunde en Informatica National research institute for mathematics and computer science in the Netherlands

24. Cont. Perfsuite (NSCA * ) * National Center for Supercomputing Applications (NCSA) Control hardware counters Measure 60 event types for the results Hardware counters

25. Stall time components on Itanium II

26. Results analysis Part one: DRDB Characterization workload on Itanium II OLTP OLAP Part two: MMDB issue Characterization of MMDB TPC-H workload --------------------------------------------------------------------------------------------------------- Dawei liu, Shan Wang, Biao Qin, Weiwei Gong: Characterizing DSS Workloads from the Processor Perspective. The International Workshop on Database Management and Application over Network DBMAN 2007: 235-240 DaweiDawei Liu, Shan Wang, Qiming Chen, Yun Tian, Weiwei Gong “Main Memory Database TPC-H Workload Characterization on Modern Processor,” Renmin University of China., TR-01, 2007, http://deke.ruc.edu.cn/tr/TR 2007-01.

27. Memory stall time breakdown TPC-H Workload on a DRDB

28. Index Influence TPC-H Workload on a DRDB

29. Branch Instruction Misprediction TPC-H Workload on a DRDB

30. DRDB vs. MMDB

31. Storage Architecture Influence

32. Summary Characterized workload on Itanium II based platform; Characterized MMDB read optimized workload on modern processors; Compare the workload breakdown of DRDB and MMDB; Explored the difference of column-oriented and row-oriented storage models in CPU and cache utilization; Investigated the index influence at low level

33. Outlines 1. Background 2. Motivation 3. Our research 4. Future works

34. Future works In-depth analysis of the results Develop new parallel techniques Instruction level parallelism MMDB benchmark issue The results expected to benefit The performance optimization of DBMS; The architecture of next-generation memory-oriented databases.

35. The End Thanks! Welcome to visit RUC. | Dawei Liu | School of Information | Renmin University of China | liudawei@ruc.edu.cn | | http://deke.ruc.edu.cn | | Tel.: +86 (10) 62513934 | Key Laboratory of Data Engineering and Knowledge Engineering MOE