Database-Presentatio.. - Department of Electrical Engineering

advertisement
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of
small Database Systems
Faculty:
Course:
Instructor:
Computer Science and Engineering, York University (CA)
CSE 6421- Advanced Database Systems
Prof. Jarek Gryz
Speaker:
Date:
Benedikt Iltisberger
2010-12-01
Hochschule
Bonn-Rhein-Sieg

Presentation Time:


Energy Consumption of DB Systems
35-45 minutes
Questions:

I am looking forward to answer your
questions during the presentation.
Slide 2
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Agenda of the Presentation

What is the topic?

Why is it important?

How to solve the problem?

What do you have to keep in mind?

Does it work in the real world?

What is the conclusion?

What can be done in the future?
Slide 3
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
What is the Topic?

Show a way how to measure the power
consumption of a single query!
Speed
Energy
Consumption
Slide 4
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
What are the Scientific Challenges?

Never done before[12]

Could/should be implemented in any DBMS

Multiple factors need to be considered

Upcoming and important research area
Slide 5
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Energy Consumption
Source: [7]
Slide 6
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Why is it worth to focus on
energy saving?

Saving the environment

Growing market  Job opportunity

Saving costs for companies

Energy bills are rising quickly

Very important research area in Europe

Interesting from the technical point of view
Slide 7
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Benchmarking Organizations

SPEC


“New SPECweb2009 benchmark adds ability
to measure power consumption of web
servers” (June 5, 2009)[4]
TPC

“Transaction Processing Performance Council
Announces First Results for its TPC-Energy
Specification” (July 7, 2010)[5]
Slide 8
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Energy consumption in the IT
Source: [6]
Slide 9
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
What have DBs to do with that?

Core Service

Delivering data for any business process

Store the foundation of a company

Need to be accessible 24/7

Today integrated even in small projects

Data warehouses and Data marts

Long running energy consuming queries
Slide 10
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Reasons for Decision-Makers
1.
Reducing costs
2.
Keep up a good image (Green IT)
1.
Innovation
2.
Keep up with the times
3.
Being faster than competitors
3.
Depending on the area of work
4.
Additional company policies
Slide 11
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Idea of this Proposal

Estimate the energy consumption of a
single query to





Compare queries in terms of efficiency
Optimize queries for energy saving
Get energy information in real time
Teach the optimizer to do this task
Green IT initiative
Slide 12
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
How to achieve this goal?

DBMS (e.g. DB2) can generate reports
about executed queries containing:


Original statement -> optimized statement
Access Plan with many details like:



CPU Cost
 CPU
I/O Cost
 HDD
Bufferpool Buffers  Main Memory
Slide 13
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
What information is given?

IBM DATABASE 2 Explain Table Format Tool

DB2 version: 09.07.3

Detailed information about queries including

Access Plan (also graphical)

Original query  Optimized Query

Detailed information about the different costs

Detailed information about tables, indexes, …
Slide 14
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Information from the Statistic File
Database Context:
---------------Parallelism:
CPU Speed:
Comm Speed:
Buffer Pool size:
Sort Heap size:
Database Heap size:
Lock List size:
Maximum Lock List:
Average Applications:
Locks Available:
None
1.889377*107
100
6000
1000
900
4096
22
1
28835
4) SORT : (Sort)
---------------Cumulative Total Cost:
Cumulative CPU Cost:
Cumulative I/O Cost:
Cumulative Re-Total Cost:
Cumulative Re-CPU Cost:
Cumulative Re-I/O Cost:
Cumulative First Row Cost:
Estimated Bufferpool Buffers:
27541.2
5.39421*109
23667.2
27541.1
5.39388*109
0
27541.2
585
Detailed information: [9, 10, 11]
Slide 15
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Key Metrics in Detail

CPU Cost


I/O Cost


Number of seeks and page transfers
Bufferpool Buffer


Instructions needed to execute a query
Number of used pages
Total Cost

Measured in Timerons (proprietary IBM algo.)
Slide 16
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
The Optimizer’s Access Plan
Slide 17
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Slide 18
Hochschule
Bonn-Rhein-Sieg
|
18000
TABLE: TPCDS
ITEM
Q2
Energy Consumption of DB Systems
/------------------+------------------\
211.307
299.933
^HSJOIN
FETCH
(
8)
( 18)
26913.4
30.4522
23078.2
4
/----------+-----------\
/---+----\
13889.9
29221.1
300
300
NLJOIN
TBSCAN
IXSCAN
TABLE:
TPCDS
(
9)
(
17)
(
19)
PROMOTION
17710.9
9200.94
14832.2
8246
/-------+--------\
|
352.257
39.4311
1.9208e+006
TBSCAN
FETCH
TABLE: TPCDS
( 10)
( 13)
CUSTOMER_DEMOGRAPHICS
723.246
39.5703
Q4
656
32.9996
|
/----+-----\
352.257
39.4311
2.8804e+006
SORT
RIDSCN
DP-TABLE: TPCDS
( 11)
( 14)
STORE_SALES
723.246
7.59231
Q5
656
1
|
|
352.257
39.4311
0.105098
0
|
300
INDEX: SYSIBM
SQL100701103316930
Q1
Q1
Slide 19
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Parse and Compute
1.
Perl script to gather important information


2.
Provide script with following parameters:



3.
Grab all CPU, HDD and Memory operations
Sum them up
Energy consumption of 1 CPU cylce
Energy consumption of 1 HDD operation
Energy consumption of 1 Memory operation
Calculate the query's power consumption

Using the information from no. 1 and no. 2
Slide 20
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Perl Script
Slide 21
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Assumed DB System

CPU[1]



HDD[2]



Overall:
213.66W
Seagate Cheetah 15K, 600GB, 16MB Cache
I/O Data transfer rate 600MB/s, 16.35W (avg operating power
consumption)
RAM[6]


Intel Xeon X7560 @2.266 GHz, 130W
8 Cores, 16 Threads, 24MB L3 Cache
4GB DDR3 RAM (2 * 2GB DIMM#) with 9W per DIMM  18W
Additional Hardware[6]:

30% of average power consumption of other components which
leads to (130W + 16.35W + (2 * 9W)) * 0.3 = 49,305W
Source: [6], Page 1233, Section 3.2
#) DIMM  Dual In-Line Memory Module
Slide 22
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
How to measure the parameters? I

CPU



Peak power consumption (p) of the manufacturer (Watts
per hour)
Total MIPS per hour (c)
Calculate avg power consumption (a)


p/c=a
Problems


Slide calculation flaws may end in high differences
Many instructions take more the one cycle


This also depends on the CPU type
Current multicore architecture need special customization
Slide 23
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
How to measure the parameters? II

HDD



Peak power consumption (p) of the manufacturer
(Watts per hour)
Calculate the max. avg I/O operations per hour (c)
Calculate avg power consumption per I/O and hour (a)


p/c=a
Problems


Different energy consumption for read & write
RAID-n or other storage systems
Slide 24
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
How to measure the parameters? III

Main Memory



Peak power consumption (p) of the manufacturer
(Watts per hour)
Calculate the max. page change rate per hour (c)
Calculate avg power consumption per page change (a)


p/c=a
Problems:


Very complicated and inprecise
Strongly dependent on memory architecture and
manufacturing process
Slide 25
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Exemplary Technical Issue:
“Samsung DDR3”
Source: [8]
Slide 26
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Automation of Data Collection
about CPU, HDD, RAM

Proposal: “energyDB”

Name, Modell, Speed, Energy consumption, …

Online service to provide needed information

Comparable with CDDB[13] or freedb[14]

Useful for other projects as well
Slide 27
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Optimizing Queries

Areas of optimization:

Execution speed (typical parameter)


Resource usage


When time is relevant
For many concurrent users/queries
New: Energy Consumption
Slide 28
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Optimizing Queries No. 1

Original Query:
SELECT i_item_id,
CAST(avg(CAST(ss_quantity AS DECFLOAT)) AS DECIMAL(10,6))
agg1, CAST(avg(ss_list_price) AS DECIMAL(10,6)) agg2,
CAST(avg(ss_coupon_amt) AS DECIMAL(10,6)) agg3,
CAST(avg(ss_sales_price) AS DECIMAL(10,6)) agg4
FROM tpcds.store_sales, tpcds.customer_demographics, tpcds.date_dim,
tpcds.item, tpcds.promotion
WHERE ss_sold_date_sk = d_date_sk and ss_item_sk = i_item_sk and ss_cdemo_sk
= cd_demo_sk and ss_promo_sk = p_promo_sk and cd_gender = 'M' and
cd_marital_status = 'S' and cd_education_status = 'College' and
(p_channel_email = 'N' or p_channel_event = 'N') and d_year = 2000
GROUP BY i_item_id
ORDER BY i_item_id
FETCH FIRST 100 ROWS ONLY
Slide 29
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Optimizing Queries (Results)
Original Query 1
(with index)
CPU Cost
I/O Cost
Buffer Pool usage*
Energy Consumption
Total Cost
Declined Query 1
(index deleted)
4.96336*109
5.39446*109
8636.78
23667.20
68303.99
70816.80
TBD
TBD
12119.3
27541.2
*) Sum of all buffer pool operations for all involved queries.
Slide 30
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Optimizing Queries No. 2

Original Query:
SELECT i_item_id,
CAST(avg(CAST(ss_quantity AS DECFLOAT)) AS DECIMAL(10,6)) agg1,
CAST(avg(ss_list_price) AS DECIMAL(10,6)) agg2,
CAST(avg(ss_coupon_amt) AS DECIMAL(10,6)) agg3,
CAST(avg(ss_sales_price) AS DECIMAL(10,6)) agg4
FROM tpcdsnp.store_sales, tpcdsnp.customer_demographics, tpcdsnp.date_dim,
tpcdsnp.item, tpcdsnp.promotion
WHERE ss_sold_date_sk = d_date_sk and ss_item_sk = i_item_sk and
ss_cdemo_sk = cd_demo_sk and ss_promo_sk = p_promo_sk and cd_gender = 'M’
and cd_marital_status = 'S' and cd_education_status = 'College' and
(p_channel_email = 'N' or p_channel_event = 'N') and d_year = 2000
GROUP BY i_item_id
ORDER BY i_item_id
FETCH FIRST100 ROWS ONLY
Slide 31
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Optimizing Queries (Results)
Original Query 2
(with index)
CPU Cost
I/O Cost
Buffer Pool usage*
Energy Consumption
Total Cost
Declined Query 2
(index deleted)
4.51373*107
1.45693*109
654.55
8495.00
3971.55
17575.20
TBD
TBD
1124.43
8772.47
*) Sum of all buffer pool operations for all involved queries.
Slide 32
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Optimizing Queries (Problems)

Better performance but higher energy
usage:



Precomputation for faster response times
Materialized Views
CPU


Cheap in terms of time consumption but
Expensive in terms of energy consumption
Slide 33
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Conclusions I

Reducing costs is always welcome in companies

High need for this research in the long term

Due to corporate identity and rising energy costs

Environmental saving

Query optimization does not necessarily
lead to power saving

Proposed technique can be added to the optimizer
Slide 34
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Conclusions II

Cons

You need to measure the single components in the
first place

Not easily applicable to complex systems

Accuracy must be determined

Checking results with hardware based measurement data
How to measure VMs?
Slide 35
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Future Work I

Add power consumption optimization techniques
directly to the optimizer

IBM seems to be working on that[12]

Optional output of estimate power consumption

Improve calculation techniques


Even slide inaccuracy can have a big impact

RAM usage is pretty inaccurate at the moment
Automization of the process
Slide 36
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Future Work II

Implementation of complex Dataware House
systems or for distributed database systems

Inaccuracy with measuring multi CPU systems

„energyDB“ web service implementation

Impact of hardware power saving

GPU for databases, SSD, Dynamic voltage scaling, ...
Slide 37
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Concluding …
This topic is worth for further research!
Slide 38
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Thank you for your attention!
Any questions?
Reference:
1.
2.
3.
4.
5.
6.
http://www.intel.com/p/en_US/products/server/processor/xeon7000/specificat
ions
http://www.seagate.com/www/en-us/products/enterprise-harddrives/cheetah-15k#tTabContentSpecifications
http://www.gruene-it.de/index.php/2010/09/17/ram-stromverbrauch-server/
http://www.spec.org/web2009/press/release.html
http://www.tpc.org/information/press/tpcpress20100707.asp
M. Poess, R. O. Nambiar, Energy Cost, “The Key Challenge of Today’s Data
Centers: A Power Consumption Analysis of TPC-C results”,Oracle
Corporation and Hewlett-Packard Company, N.A..
Slide 39
Hochschule
Bonn-Rhein-Sieg
Energy Consumption of DB Systems
Reference (cont.):
7.
8.
9.
10.
11.
12.
13.
14.
15.
http://www.bp.com/liveassets/bp_internet/globalbp/globalbp_uk_english/repo
rts_and_publications/statistical_energy_review_2008/STAGING/local_assets
/2010_downloads/statistical_review_of_world_energy_full_report_2010.pdf
http://www.samsung.com/global/business/semiconductor/Greenmemory/Appl
ications/ServerStorage/ServerStorage_DDR3.html
http://www-01.ibm.com/support/docview.wss?rs=0&uid=swg21207055
http://publib.boulder.ibm.com/infocenter/db2luw/v8/index.jsp?topic=/com.ibm.
db2.udb.doc/admin/r0000295.htm
http://www.ibm.com/developerworks/data/library/techarticle/0212wieser/index
.html
http://www.freepatentsonline.com/y2009/0281986.html
http://www.gracenote.com/
http://www.freedb.org
R. Ramahrishnan, J. Gehrke, „Database Management Systems”, Third
Edition, McGraw-Hill, Boston, 2003.
Slide 40
Download