Transcript ppt

Towards Eco-friendly
Database Management
Systems
W. Lang, J. M. Patel (U Wisconsin),
CIDR 2009
Shimin Chen
Big Data Reading Group
Introduction

Energy consumption is important for data
centers:
 2005:
1.2% of total US energy consumption is
attributed to powering and cooling servers, ~ $2.7B
 If current methods for powering data centers
continue, the consumption will nearly double by 2011

For DBMS:
 Previously
large ignored energy efficiency
 Must start considering energy as a critical metric
This paper: ecoDB


New project: energy efficient data processing
techniques
Two broad classes of techniques:
 “global”:

change how entire system is managed or used
E.g. job scheduling
 “local”:
improve methods of processing data at
individual nodes (focus of the paper)
Idea
Two Questions

(1) “How does a system generate graphs as
shown in Figure 1?”
 DMBS
must know HW capabilities and operating
characteristics
 Accurately estimate / continuously measure energy
consumption

(2) “How can such a graph be used?”
 Systematic
method to change settings
 Service level agreements (SLAs)

This paper focuses on mechanisms for creating
graphs
Outline





Introduction
Processor Voltage/Frequency Control (PVC)
Improved Query Energy-efficiency by
Introducing Explicit Delays (QED)
Opportunities for Energy Efficiency
Summary
Techniques

CPU freq = front side bus (FSB) freq * CPU
multiplier

DVFS (dynamic voltage and frequency scaling)
 Each
p-state defines a CPU multiplier
 CPU voltage is based on CPU multiplier

Under-clocking (Focus of this paper)
 Reduce
FSB freq
 Finer granularity
 Also changes RAM freq
System Under Test

System components:









ASUS P5Q3 Deluxe Wifi-AP motherboard
Intel Core2-Duo E8500
2×1GKingston DDR3 main memory
ASUS GeForce 8400GS 256M
Western Digital Caviar SE16 320G SATA disk
Power supply unit (PSU): a Corsair VX450W PSU
System power draw measured by a Yokogawa WT210
unit (suggested by SPEC Power benchmark)
MS Windows Server 2008
JDBC (Java 1.6)
Power

CPU power sensors on motherboard:
 ASUS
motherboard has an EPU processor that
directly measures the CPU power.
 ASUS P5Q3 Deluxe 6-Engine software displays
information gathered from this hardware sensor.

Current CPU wattage displayed in GUI:
 The
authors sample the GUI every second
 Compute CPU joules using the average CPU wattage
and the execution time of a workload
Component powers


No hard disk, no operating system
Focusing on CPU power:

CPU power consumption is often about 25% of the total system
power consumption in the experiments
DB test

Workload





FSB underclocking (allowed by ASUS 6-engine software)



Stock (normal)
Reduce FSB freq by 5%, 10%, and 15%
CPU voltage downgrade


Use a commercial DBMS and MySQL 5.1.28
TPC-H (ad-hoc decision support), scale factor 1.0 (1GB data)
Only run Query 5: six table join and a group by
A run consists of ten queries with different parameters
“small” and “medium” downgrade
7 settings:

Stock, 3 FSB freq reductions X 2 CPU voltage downgrades
Equal
Energy
delay
product
Equal
Energy
delay
product
With the same voltage level, larger frequency the better EDP
Theoretical Modeling



EDP= joules x times = power x time2
= power / freq2
Power=CV2F
EDP = CV2/F
Disk Energy


Measured separately for stock setting
Warm database
 CPU:
1228.7 Joules
 Disk: 214.7 Joules

Cold database
 CPU:
2146.0 Joules
 Disk: 1135.4 Joules
Outline





Introduction
Processor Voltage/Frequency Control (PVC)
Improved Query Energy-efficiency by
Introducing Explicit Delays (QED)
Opportunities for Energy Efficiency
Summary
Idea




Explicitly delay queries
look for commonalities among multiple queries
Group multiple queries into a single query
After execution, split query results
Setting

DB clients repeatedly issue single table select
queries with different selection predicate.
 For
example:
SELECT *
FROM lineitem
WHERE l_quantity=X



DBMS processes one query at a time
QED: buffer queries in a queue, merge them,
send the merged query, split results
In the experiments, X is different for the queries,
so no overlaps
As batch size increases, diminishing decrease in energy
consumption.
Outline





Introduction
Processor Voltage/Frequency Control (PVC)
Improved Query Energy-efficiency by
Introducing Explicit Delays (QED)
Opportunities for Energy Efficiency
Summary
Opportunities in (DBMS) Software


Traditional DB investigations into improving
query response times
Energy vs. performance tradeoffs
 Operator-level:
rethink join algorithms
 Query-level: energy-efficient query plans
 Workload management per server
 Workload management for the entire collection of
servers: scheduling and using techniques to turn
entire servers off
Summary


Energy-efficient data processing
Studied two techniques
 Processor
Voltage/Frequency Control (PVC)
 Improved Query Energy-efficiency by Introducing
Explicit Delays (QED)

Designing a DBMS to balance the response time
vs. energy consumption opens a wide range of
research issues