Transcript Poster - Advanced Data Management Technologies Laboratory

NSF Grant Number: IIS-0133686
Institution: Carnegie Mellon University
PI: Anastassia Ailamaki
Title: CAREER:
Bridging Databases and Computer Architecture: Optimizing DBMS for Deep Memory Hierarchies
Research Objectives
Significant Results
To bridge database research to computer architecture by
• The PAX cache-aware data placement yields a 2x speedup
optimizing database workload performance when running on
for decision-support queries (VLDBJ’02).
modern CPUs with deep memory hierarchies.
• DB microbenchmarks to measure micro-architectural
behavior (CMU TR’03).
Approach
• Workload characterization combined with machinelearning leads to predictable performance (MASCOTS’04).
Eliminate unnecessary memory references, by increasing
spatial and temporal locality using the following techniques: • By Synchronizing Transactions Through Explicit Processor
Scheduling we achieve 40% TPC-C speedup (VLDB’04).
• Static data placement for deep memory hierarchies using
three optimizing low-cost, high-impact data layouts.
Fates: Automatically-Tuned Database Storage (VLDB’04)
• Smart instruction reorganization techniques for high
utilization and lower miss ratio in instruction caches.
cache
Operators
• Cache Buffer Management to overcome the static cache
memory
page hdr Buffer
Clotho
configuration and replacement policy.
pool
• Reducing coherence overhead on multiprocessor
Bufferpool
Manager
systems using sharing classification.
disks
Lachesis
Broader Impact
Storage Manager
Education: Several graduate and undergraduate students
are building valuable interdisciplinary background in
databases and computer architecture.
Society: Half the processors in high-end servers run
database systems, so database performance is crucial for
commercial applications.
Technology transfer: to industry, via conference talks and
direct contacts with IBM, Oracle, and Intel
Connections: to both database and architecture research
DBMS use same data layout
at all memory levels
Clotho: new in–memory page layout
Atropos
disk 0
disk 1
Logical Volume Manager
• query-specific organization
• decoupled from storage!
Lachesis: efficient I/O construction
• explicit hints
• restored optimizer assumptions!
Atropos: new disk page layout
• exploits unique disk features
• transparent device mgmt
data directly placed via
scatter/gather I/O
disk array
Fates:
Highest memory utilization
At all levels
Under any workload
Papers: VLDB’03, FAST’04, VLDB’04