Evaluating Energy and Performance for Server-
Class Hardware Configurations
Chenguang Liu, Jianzhong Huang, Qiang Cao, Shenggang Wan, Changsheng Xie
School of Computer Science and Technology, Huazhong University of
Science and Technology, Wuhan, China
Challenge Problems
• Data Centers 15-40x the energy intensity of typical office buildings
• A single rack of servers can be 20 kW – $17k per year (at $.10/kWh) per rack– Hundreds of racks per center
• For high-end data centers, energy cost increased as much as 25% annually, and making it a big consideration in the total cost of ownership.
Problem Example
• Strategies (software & hardware)—migrate workloads & shut down—become more energy proportional
• All of them need qualitative and quantitative reference
Energy & Performance Metrics
• Energy—the average power(Watts)
• Performance—I/O operations per second(IOPS)
• Integrated factor—I/O operations per joule
Workload Categories cont.
• FileServer—a file system workload similar to SPECsfs
• Varmail—a /var/mail NFS mail server emulation
• WebServer—a mix of open/read/close ops of multiple files in a directory tree
• OLTP—a database emulator using I/O model from Oracle 9i
Workload Properties
Workload file mail web database
Average file size
64MB 16KB 64KB 10MB
Average dir depth
500 1 50 1
Number of files
10000 1000 1000 10
I/O sizes R/W
1MB 1MB 1MB 2KB
Number of threads
50 16 100 200
R/W Ratio 1:2 1:1 10:1 1:1
Test BedSpecification Machine InformationServer model Dell R710Machine age 1 yrCPU model Intel Xeon 5500CPU speed 2.26GHz
No. of CPUs 2 dual core(adjustable)L1 cache size 16KL2 cache size 2ML3 cache size 8M
RAM size 4*4G DDR3 (adjustable)RAM type DIMMDisk Type SASDisk Size 137GDisk RPM 7200Disk cache 8M
Chipset Intel 5520
Elements of Total Energy Consumption
—Woverhead =energy for maintaining the server's fundamental operating mode
—WCPU, Wmem & Wdisk =energy for workloads on different hardware parts respectively
—α, β & γ=the weight factor
CPU mem disk overheadW W W W W
Average Power in 6 configurations
1u4g 1u8g 1u16g 2u4g 2u8g 2u16g
Ave
ra
ge
P
ow
er(W
atts)
0
20
40
60
80
100
120
140
160
fileserver varmail webserver oltp null
• When a server does not perform any work, it consumes the most energy.
• The workload selection alone cannot reduce idle power, but combined with right-sizing techniques, it can improve power efficiency by prolonging idle periods.
• Different workloads exercise the system’s resources differently, directly affecting the additional power
1u4g 1u8g 1u16g 2u4g 2u8g 2u16g
Perf
orm
ance
(ops/
sec)
0
500
1000
1500
2000
2500
3000
1u4g 1u8g 1u16g 2u4g 2u8g 2u16gE
nerg
y E
ffic
iency
(iops/
J)0
5
10
15
20
25
30
Results in FileServer Workload
• Performance of fileserver workload is both CPU and memory-critical
• The configuration of 1u8g proved to be the most power-efficient one, achieving the top point of 26.7373(iops/J)
• Energy of fileserver workload is memory-critical
Results in Fileserver Workload(cont.)
1u4g 1u8g 1u16g 2u4g 2u8g 2u16g
Pe
rfo
rma
nce
(io
ps/s
ec)
0
200
400
600
800
1000
1200
1400
1600
1800
1u4g 1u8g 1u16g 2u4g 2u8g 2u16gE
nerg
y E
ffic
iency(
iops/J
)0
2
4
6
8
10
12
14
16
Results in Varmail Workload
• Performance of Varmail workload is CPU-critical
• The configuration of 2u4g proved to be the most power-efficient one, achieving the top point of 15.0094(iops/J)
• Energy of Varmail workload is also CPU-critical
Results in Varmail Workload(cont.)
1u4g 1u8g 1u16g 2u4g 2u8g 2u16g
Pe
rfo
rma
nce
(io
ps/s
ec)
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
1u4g 1u8g 1u16g 2u4g 2u8g 2u16gE
ne
rgy
Eff
icie
ncy
(io
ps/
J)
0
20
40
60
80
100
120
140
160
180
200
Results in WebServer Workload
• Performance of webserver workload is neither CPU nor memory-critical
• The configuration of 1u4g(the least power consumption) proved to be the most power-efficient one
• Web server’s demand for hardware is easy to meet
Results in Webserver Workload(cont.)
1u4g 1u8g 1u16g 2u4g 2u8g 2u16g
Perf
orm
ance
(iops/
sec)
0
200
400
600
800
1000
1u4g 1u8g 1u16g 2u4g 2u8g 2u16gE
nerg
y E
ffic
iency(iops/J
)
0
2
4
6
8
10
Results in OLTP Workload
Results in OLTP Workload(cont.)
• Performance of database server workload is neither CPU nor memory-critical
• Like web server, 1u4g will be a good choice
In fact and intuitively, as to the characteristic of OLTP processing, α, β and γ should be great values
A reasonable explanation may come from the inner mechanism(database lock) of OLTP applications blocking the contributions of hardware enhancement.
CPU mem disk overheadW W W W W
Conclusion
• Strengthening every aspect of a server configuration is not always a wise thing(not just the matter of hardware configuration)
• Since the computing and storage requirements of web server applications are easy to meet and we can't easily improve the performance of OLTP applications, we can aggregate different workloads on the single server according to the actual situation
Conclusion (cont.)
• Even in single machine different workloads need different hardware configurations to achieve performance- and energy- efficiency
• In the management of a modern datacenter, a hardware configuration adaptor would play a very significant role