Embed Size (px)
Citation preview
By Derrick Lee, Garrett Lord, Jesse Martinez, and Christopher Moore
Mentors: HB Chen, Parks Fields, Benjamin McClelland, Alfred Torrez
Computer Systems, Cluster, and Networking Summer Institute 2011
Performance Studies of Out-of-Core Computing on Various Storage Devices
– Fusion-IO’s IOSan, SuperTalent’s
RaidDrive, OCZ Vertex 3, Kingston V+
Outline � Out of Core Testing
� Introduction
� Devices Used � Types of Benchmarks
� Results from Benchmarks � Cost Efficiency � Areas for Improvement
� Conclusion
Introduction
• Datasets are characterized by their very large sizes with multiple superposed scalar and vector fields, demanding an imperative need for new interactive exploratory visualization capabilities
• Sequence of Out-of-core benchmarking tests were done on various storage devices such as SATA based Solid State Devices and PCI-Express based SSD
Devices • System – Fedora 13 64bit – RAM: 144GB, 16 Cores – Devices • Fusion IO IOSAN SSD 640GB
• System – Fedora 15 64bit – RAM: 8GB, 8 Cores – Devices • SuperTalent RAID Drive 512GB • APTEC PCI-Express 4-Port SATA Controller – 4x 240 GB OCZ Vertex 3
– 4x 256 GB Kingston V+ Series
Benchmark Tests � FIO (File Input/Output) Benchmark
� Unbuffered/Buffered I/O � Sequential Read
� Sequential Write
� Random Read
� Random Write
� Graph 500 � Scale 28 � Cores: 16
FIO Benchmark � Three types of results done
� Aggregate Bandwidth
� I/O Operations per Second � Completion Latency
� Ext2 File systems were created on the devices
0
20
40
60
80
100
120
140
Minimum TEPS Median TEPS Maximum TEPS
Trav
erse
d E
dge
s per
Sec
ond (
Mil
lion
s)
Range of TEPS
Graph 500 Results
Memory
Fusion IO
Fusi
on I
O
Super
Tal
ent
OC
Z
Kin
gsto
n
Fusi
on I
O
Super
Tal
ent
OC
Z
Kin
gsto
n
Fusi
on I
O
Super
Tal
ent
OC
Z
Kin
gsto
n
Fusi
on I
O
Super
Tal
ent
OC
Z
Kin
gsto
n
IOP
S
FIO Benchmark - IOPS - Unbuffered - 64 Jobs
512 2k 8k 32k 128k 512k 2M 8M
Rand. Read Rand. Write Seq. Read Seq. Write
140k
120k
100k
80k
60k
40k
20k
0
0
200
400
600
800
1000
1200
1400
1600 Fu
sion
IO
Super
Tal
ent
OC
Z
Kin
gsto
n
Fusi
on I
O
Super
Tal
ent
OC
Z
Kin
gsto
n
Fusi
on I
O
Super
Tal
ent
OC
Z
Kin
gsto
n
Fusi
on I
O
Super
Tal
ent
OC
Z
Kin
gsto
n A
ggr
egat
e B
andw
idth
(M
B/s
) FIO Benchmark - Aggregate Bandwidth - Unbuffered - 64 Jobs
512 2k 8k 32k 128k 512k 2M 8M
Rand. Read Rand. Write Seq. Read Seq. Write
FIO Benchmark Cost Efficiency
Device (MTBF)
Capacity (GB) / $
Bandwidth (MB) / $
IOPS / $
Fusion IO 0.040 0.012 3.011
Super Talent (1.5 mil. Hours)
0.256 0.056 14.355
OCZ Vertex 3 (2 mil. Hours ea.)
0.426 0.035 8.904
Kingston V+ (1 mil. Hours ea.)
0.499 0.021 7.224
Cost/Capacity List
16,000
4,000 2,500 2,400
0
2,000
4,000
6,000
8,000
10,000
12,000
14,000
16,000
18,000
Fusion IOSSAN(640GB)
Supertalent RaidDrive(920GB)
OCZ Vertex3 X 4 + controller(960GB)
Kingston V+NowSSD X4 +
Controller (1024GB)
Cost($)
Bandwidth Cost Comparison
0.012
0.056
0.035
0.021
0
0.01
0.02
0.03
0.04
0.05
0.06
FusionIO IO SAN
Supertalent RaidDrive
OCZ Vertex 3 Kingston V+ NowSSD
MB/sec/$
MB/sec/$
IOPs Cost Comparison - IOPs/$
3.011
14.355
8.904
7.224
0
2
4
6
8
10
12
14
16
FusionIO IO SAN
Supertalent RaidDrive
OCZ Vertex 3 Kingston V+ NowSSD
IOPS/$
IOPS/$
Areas for Future Improvement • Benchmarks
• IOZone
• XDD
• Bonnie++
• Bandwidth+
• STREAM
• ScaLAPACK
• Performance Analysis of SSDs in Larger Systems
• Memory Swap Performance
• File Systems Formats
Conclusion • In respect to processing power, FIO test results also showed that the Fusion IO drive displayed significantly better performance. However, performance per dollar indicates that the smaller SSDs had a higher cost efficiency rating. • Graph500 benchmark testing showed that the Fusion IO had a high edge traverse rate with the swap system but baseline results showed to be higher. • Although the Fusion IO had higher performances, the performance/cost results showed to be lower than the rest of the SSD devices.
Questions?
