DBMSs On A Modern Processor: Where Does Time Go

1
DBMSs On A Modern ProcessorWhere Does Time Go?
VLDB 1999, Edinburgh, Scotland

• Anastassia Ailamaki
• David J. DeWitt
• Mark D. Hill
• David A. Wood

Brief Recap by Ippokratis Pandis
15-823 Hot Topics in DB Systems
2
New bottleneck in DBMSs 1

• Processor speeds increase faster than memory
  speeds
• DBMSs becoming compute and memory bound
• New architectures hide I/O latency
• Larger memories
• Pipelining/Out of order execution/Non-blocking
  caches
• Complexity of DBMSs
• Discover performance bottlenecks

3
New bottleneck in DBMSs 2

• Time a query needs TQTCTMTBTR-TOVL
• Memory stalls TMTL1ITL1DTL2DTL2ITDTLBTITLB

4
New bottleneck in DBMSs 2

• Time a query needs TQTCTMTBTR-TOVL
• Memory stalls TMTL1ITL1DTL2DTL2ITDTLBTITLB
• Memory stalls
• Stalls due to L2 Data cache misses
• Stalls due to L1 Instruction cache misses
• Other stalls
• Stalls due to Misprediction Branches
• Resource-related stalls

5
Outline

• Motivation
• Results
• Discussion

6
Methodology

• Four commercial DBMSs
• In the same platform
• Windows NT4.0
• Pentium II Xeon/MT Workstation
• Minimal experimental setup
• Memory resident DB
• Exclusion of I/O subsystem
• Processor hardware counters for the experiments
  (faster, little overhead, and no simulation)
• Simple queries and not whole benchmarks
• Sequential range selection
• Indexed range selection
• Sequential join

7
Results Execution Time

• Actual computation time is almost less than 50
• The major bottleneck are the memory stalls

8
Results Memory Stalls

• L1I and L2D dominate

9
Results Comparison with TPC

• Simple micro-benchmarks present similar behavior
  with the more complex TPC benchmarks

10
Summary

• DBMSs do not exploit architecture features of
  modern processors
• Developers of DBMSs should
• Optimize Data Placement for L2 (L2D)
• Optimize Instruction Placement for L1 (L1I)
• Not expect significant performance improvement
  without addressing other issues
• Branch mispredictions
• Resource-related stalls
• Using simple queries rather than full scale
  benchmarks provides an easy and accurate way to
  study the behavior of DBMSs

11
Outline

• Motivation
• Results
• Discussion

12
Critique

• What is happening in other OSs? (Linux)
• What is happening in other platforms? (AMD)
• Do we observe the same behavior?
• More complete study using I/Os and/or TPC
  benchmarks

13
Points of Discussion Vijay

• What would be the level of complication of study
  if I/O were considered?
• To what extent can the concepts proposed be
  generalized to other Operating systems and
  processors without major change?
• How can branch prediction be made more accurate?

14
Points of Discussion Ippokratis

• Cannot always reduce stalls by increasing cache
  size
• This paper presents a study that was conducted in
  1999. Now, six years later do the presented
  numbers still hold?
• New technologies in architecture and their
  affect. Multi-core, and multi-processor machines.
  How about new memory designs?
• Study with the presence of I/Os
• Is there any work in misprediction stalls and
  resource stalls related to DBMSs?

15
Thank you!!