Tech-Forward - Achieving Business Readiness For Copilot in Microsoft 365
Damon2011 preview
1. DaMoN 2011 Paper Preview
Organized by
Stavros Harizopoulos and Qiong Luo
Athens, Greece
Jun 13, 2011
2. Preview of Afternoon Program
• 13:00-15:00 Paper Session I: FLASH DISKS, FPGAS,
AND SMARTPHONES
• 15:00-15:30 Coffee Break
• 15:30-17:00 Paper Session II: MODERN CPUS AND
MEMORY SYSTEMS
• 17:00-17:30 Coffee Break
• 17:30-18:30 Panel: WHITHER HARDWARE-
SOFTWARE CO-DESIGN?
3. Paper Session I:
FLASH DISKS, FPGAS, AND SMARTPHONES
• Enhancing Recovery Using an SSD Buffer Pool
Extension
• Towards Highly Parallel Event Processing
through Reconfigurable Hardware
• QMD: Exploiting Flash for Energy Efficient Disk
Arrays
• A Case for Micro-Cellstores: Energy-Efficient
Data Management on Recycled Smartphones
10. A Case for Micro-Cellstores
Energy-Efficient Data Management
on Recycled Smartphones
Stavros Harizopoulos
Spiros Papadimitriou
1/3 The views contained herein are the authors' only and do not
necessarily reflect the views of Hewlett-Packard or Google
11. A Case for Micro-Cellstores
Energy-Efficient Data Management on Recycled Smartphones
>1 billion smartphones expected to become
obsolete in the next 5 years
What happens to old computers, servers,
cell phones?
Can we do better?
2/3
S. Papadimitriou
12. A Case for Micro-Cellstores
Energy-Efficient Data Management on Recycled Smartphones
Repurpose old smartphones
Power-profile characterization of
current-generation smartphone
Initial evaluation: up to 6x more
efficient (vs. other “wimpy” nodes)
on scan workloads
Motivate energy-efficient,
sustainable solutions
3/3
S. Papadimitriou
13. Paper Session II:
MODERN CPUS AND MEMORY SYSTEMS
• Scalable Aggregation on Multicore Processors
• How to Efficiently Snapshot Transactional Data:
Hardware or Software Controlled?
• Vectorization vs. Compilation in Query Execution
14. Scalable Aggregation on
Multicore Processors
Yang Ye, Kenneth Ross, Norases Vesdapunt
Columbia University
1/4 DaMoN 2011
15. Utilization Challenge
• What is the best way to use the shared/partitioned
resources for computations like aggregation?
• Issues:
• Coordination overhead of shared data structures
• Latches and/or atomic operations
• Contention
• Space overhead of replicated data structures
• With n threads, each thread gets 1/nth of the shared cache
and RAM
• Robustness under many input data distributions
2/4 DaMoN 2011
16. Niagara vs Nehalem
• Prior work on Sun Niagara T1 and T2 machines
• Some TPC benchmark winners use the T2 (!)
• Many threads: high parallelism
• Do these results generalize to other architectures such as
the Nehalem processor?
• Differences in:
• Clock speed
• Relative cost of a miss
• Degree of parallelism
• Memory hierarchy & consistency model
• Core sophistication (pipelines, branch prediction, etc.)
3/4 DaMoN 2011
17. Architecture Dependence
• How architecture-independent can a high-
performance implementation be?
4/4 DaMoN 2011
18.
19.
20. Vectorization vs.
Compilation
in Query Execution
Juliusz Sompolski
Peter Boncz
June 13th, 2011
DaMoN 2011, Athens, Greece Marcin Zukowski
1/3
21. Vectorization vs.
Compilation:
get rid of interpretation
overhead
Vectorization processes data in blocks to
amortize interpretation overhead over
multiple tuples.
JIT query compilation generates and
compiles specialized program for each
query remove interpretation at all.
Both get rid of interpretation overhead.
2/3
22. Vectorization vs.
Compilation
Once we’re rid with interpretation
overhead... are they worth combining?
Vectorized systems could use compilation to
move to tuple-at-a-time processing without
interpretation overhead in some operations.
Existing systems using JIT compilation still
choose to work tuple-at-a-time. Should they
sometimes switch to vector-at-a-time model?
Case studies and examples.
3/3
23. Summary
• An exciting afternoon program ahead
– Seven interesting papers in two sessions
• Flash disks, FPGAs, and (recycled) smartphones
• Modern (multicore) CPUs and memory systems
– Panel with experts on hardware-software co-
design issues