Part No 821-0125-10Revision 1.0, 06/25/09
SOLID STATE DRIVES IN HIGH PERFORMANCE COMPUTINGREDUCING THE I/O BOTTLENECKLawrence McIntosh, Systems Engineering Solutions GroupMichael Burke, Ph.D., Strategic Applications Engineering
Sun BluePrints™ Online
Sun Microsystems, Inc.
Table of ContentsIntroduction ....................................................................................................... 3
Motivation ..................................................................................................... 3
SSD technology review ........................................................................................ 4
Single system application performance ............................................................... 6
The ABAQUS benchmark application ................................................................ 7
Hardware configuration .............................................................................. 8
Software configuration ................................................................................ 8
The NASTRAN benchmark application ............................................................... 9
Hardware configuration ............................................................................ 11
Software configuration .............................................................................. 11
The ANSYS benchmark application ................................................................. 12
Hardware configuration ............................................................................ 13
Software configuration .............................................................................. 13
Summary for single system application performance ...................................... 14
SSD usage with the Lustre parallel file system .................................................... 14
Lustre file system design ............................................................................... 14
IOZone file system testing ............................................................................. 18
Hardware configuration ............................................................................ 20
Software configuration .............................................................................. 20
Summary for SSD usage with the Lustre parallel file system ............................ 20
Future directions ........................................................................................... 20
Conclusion ....................................................................................................... 21
Appendix: Benchmark descriptions and parameters ............................................ 22
ABAQUS standard benchmark test cases ......................................................... 22
Hardware configuration ............................................................................ 26
Software configuration .............................................................................. 26
NASTRAN benchmark test cases ..................................................................... 27
Hardware configuration ............................................................................ 28
Software configuration .............................................................................. 28
ANSYS 12.0 (prel. 7) with ANSYS 11.0 distributed benchmarks .......................... 29
Hardware configuration ............................................................................ 30
Software configuration .............................................................................. 30
About the authors ............................................................................................. 30
References ........................................................................................................ 31
Ordering Sun Documents................................................................................... 31
Accessing Sun Documentation Online ................................................................ 31
Sun Microsystems, Inc.3 Solid State Drives in HPC: Reducing the I/O Bottleneck
IntroductionThis Sun BluePrints™ article focuses on a comparison between traditional hard disk
drives (HDDs) and the newer solid state drive (SSD) technology in high-performance
computing (HPC) applications. SSD devices can help correct the imbalance between
processor and storage speed while also reducing energy usage and environmental
impact. This comparison was performed using two approaches:
• Application-based benchmarking was performed using the ABAQUS, NASTRAN,
and ANSYS finite-element analysis (FEA) applications, in order to evaluate the
effect of SSD technology in realistic HPC applications. These applications are
commonly used to benchmark HPC systems.
• Benchmark testing of storage performance using the Lustre™ parallel file
system and the popular IOZone benchmark application was performed, in order
to evaluate large sequential I/O operations typical for the Lustre file system
employed as a compute cluster data cache. These tests were performed using
three system configurations:
– A baseline test using the Lustre file system with a single HDD-based Object
Storage Server (OSS)
– A Lustre file system configuration using a single SSD-based OSS similar to the
baseline test
– A comparison test using the Lustre file system and two SSD-based OSSs in
parallel
The results of these tests demonstrate the potential for significant benefits in the
use of SSD devices for HPC applications with large I/O components.
MotivationProcessor performance, especially in high-performance clustered multiprocessor
systems, has grown much more quickly than the performance of I/O systems and
large-scale storage devices. At the same time, high-performance computing tasks
in particular have been dominated more and more by the need to manage and
manipulate very large data sets, such as sensor data for meteorology and climate
models. In combination, the need to manage large data sets while meeting the data
demands of fast processors has led to a growing imbalance between computation
and I/O — the I/O bottleneck.
This I/O bottleneck constrains the overall performance of HPC systems. It has
become essential to look for HPC performance improvements somewhere other than
increased processor speed. It has become equally essential to reduce the energy
requirements of HPC systems. Many HPC datacenters are up against hard limits of
available power and cooling capacity. Reducing energy cost and cooling load can
Sun Microsystems, Inc.4 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
provide increases in capacity that would otherwise not be feasible. The use of solid
state devices to replace traditional HDDs can allow HPC systems to both improve I/O
performance, and reduce energy consumption and cooling load.
This Sun BluePrints article is divided into the following sections:
• “SSD technology review” on page 4 provides an introduction to SSD technology.
• “Single system application performance” on page 6 compares HDD and SSD
technology using well-known HPC applications.
• “SSD usage with the Lustre™ parallel file system” on page 14 compares an HDD
baseline configuration with SSD-based configurations for the Lustre file system.
• The “Appendix: Benchmark descriptions and parameters” on page 22 details
specifics of benchmarks used in this study.
SSD technology reviewSSD devices are already familiar to most, in the form of flash drive technology used
in PDAs, digital cameras, mobile phones, and in USB thumb drives used for portable
storage and data transfer. With no moving parts, high speed data transfer, low power
consumption, and cool operation, SSD devices have become a popular choice to
replace HDDs.
There are two choices available for SSD technology: multilevel cell (MLC) SSDs
as found in laptops and thumb drives, and single-level cell (SLC) SSDs as used in
enterprise servers. In MLC storage, data is stored with two bits in each storage cell.
SLC storage stores a single bit per cell, so MLC devices store twice as much data as
SLC devices for the same storage footprint. SLC devices, however, are faster and have
ten times the life expectancy of MLC devices. Sun enterprise SSD devices use SLC
technology.
The experiments described in this article used the Intel® X25-E Extreme SATA Solid-
State Drive mounted in either a 3.5 inch SATA Carrier or 2.5 inch SAS, similar to that
shown in Figure 1. These SSDs deliver very good performance while simultaneously
improving system responsiveness over traditional HDDs in some of the most
demanding applications. Sun SSDs are available in both 2.5-inch and a 3.5-inch
carriers to support a wide variety of Sun rack mount and blade servers. By providing
these two formats, SSDs can be used as a drop-in replacement for HDDs, while
delivering enhanced performance, reliability, ruggedness and power savings.
Sun Microsystems, Inc.5 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
Figure 1. Solid state drive mounted in 3.5 inch SATA carrier
SSDs yield access to stored data via traditional read operations and store data via
traditional write operations. No modifications are required to applications that
access data via HDDs. SSDs are much faster and provide greater data throughput
than HDDs, because there are no rotating platters, moving heads, fragile actuators,
unnecessary spin-up time or positional seek time. The SSDs employed by Sun utilize
native SATA interface connections so they do not require any modification to the
hardware interface when placed in Sun servers.
SSDs utilize native SATA interfaces, but also provide a built-in parallel NAND channel
to the flash memory cells (Figure 2). This architecture provides much greater
performance compared to traditional HDDs without modification to applications.
SSDs also support native command queueing (NCQ), lowering latency and increasing
I/O bandwidth. Sun SSDs incorporate a wear-leveling algorithm for higher reliability
of data and provide a life expectancy of two million hours Mean Time Between
Failures (MTBF).
Sun Microsystems, Inc.6 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
SATA Interface Flash Memory
Channel ...
Flash MemoryChannel 0
Flash MemoryChannel n
IntelSystem
On a Chip(SOC)
NANDFlash
Memory
NANDFlash
Memory
NANDFlash
Memory
Figure 2. SSDs use a native SATA interface, but provide fast parallel NAND channels
Single system application performance To evaluate the use of SSDs in HPC environments, Sun first compared HPC
application run times on Sun Fire™ servers using traditional HDDs and SSDs. These
comparisons were made using FEA applications from the mechanical computer aided
engineering (MCAE) domain. These FEA applications are computer models that focus
on designs and materials used in real engineering analysis. More important for
this study, these applications are the basis for a number of well-known application
benchmarks, widely used to evaluate HPC systems. In this study, the results of
similar computations using HDD and SSD configurations were compared. If these
application benchmarks run significantly faster in the SSD configuration, then
running these applications with data from real user models might show similar
gains.
The FEA benchmark applications used in this report are:
• ABAQUS
• NASTRAN
• ANSYS
Note: For details of benchmarks and benchmark configurations, see the Appendix beginning on page 22.
Sun Microsystems, Inc.7 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
1 For details of the Sun Fire x4450 server, please see http://www.sun.com/servers/x64/x4450/
The ABAQUS benchmark applicationRuns were made on a Sun Fire X4450 server1 using from one to four cores per job
with the ABAQUS standard test suite. (Please see “ABAQUS standard benchmark test
cases” on page 22 for details.) This suite features "large" models, that require:
• Large number of degrees of freedom
• Large memory requirements
• A substantial I/O component
The Sun Fire X4450 server with four 2.93 GHz quad-core Intel Xeon® Processor X7350
CPUs demonstrated a substantial performance improvement using Sun SSDs as
compared to traditional HDDs. The performance generally increased in concert with
the system load (increased number of active cores). Table 1 illustrates the overall
comparisons for the ABAQUS standard test suite runs.
Table 1. ABAQUS Benchmark standard test suite: HDDs vs. SSDs
Test, cores Time(sec)
x4450 HDD
Time (sec)
x4450 SSD
Time Ratio
HDD:SSD
Improvement Sockets Cores
S2a-1,1 2787 2464 1.13 12.00% 1 1
S2a-2,2 1659 1298 1.28 22.00% 2 2
S2a-4,4 949 709 1.34 25.00% 4 4
S2b-1,1 3074 3111 0.99 -1.00% 1 1
S2b-2,2 1684 1753 0.96 -4.00% 2 2
S2b-4,4 1608 1606 1 0.00% 4 4
S4a-1,1 679 613 1.11 10.00% 1 1
S4a-2,2 628 419 1.5 33.00% 2 2
S4a-4,4 480 303 1.58 37.00% 4 4
S4b-1,1 11698 8115 1.44 31.00% 1 1
S4b-2,2 6162 4520 1.36 27.00% 2 2
S4b-4,4 3734 2655 1.41 29.00% 4 4
S4c-1,1 6608 6743 0.98 -2.00% 1 1
S4c-2,2 5499 4571 1.2 17.00% 2 2
S4c-4,4 4073 3509 1.16 14.00% 4 4
S5-1,1 1708 1051 1.63 38.00% 1 1
S5-2,2 1345 675 1.99 50.00% 2 2
S5-4,4 1069 456 2.34 57.00% 4 4
S6-1,1 9040 7175 1.26 21.00% 1 1
S6-2,2 6128 4741 1.29 23.00% 2 2
S6-4,4 4864 3520 1.38 28.00% 4 4
As shown in Figure 3, the use of SSD improves performance in all cases, with
increases as great as two times the HDD baseline in the S5 test.
Sun Microsystems, Inc.8 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
Figure 3.
0
2000
4000
6000
8000
10000
12000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Tim
e (s
ec)
Test - # sockets, # cores
HDD
SSD
S2a
- 1, 1
S2a
- 2, 2
S2a
- 4, 4
S2b
- 1, 1
S2b
- 2, 2
S2b
- 4, 4
S4a
- 1, 1
S4a
- 2, 2
S4a
- 4, 4
S4b
- 1, 1
S4b
- 2, 2
S4b
- 4, 4
S4c -
1, 1
S4c -
2, 2
S4c -
4, 4
S5 - 1
, 1S5
- 2, 2
S5 - 4
, 4S6
- 1, 1
S6 - 2
, 2S6
- 4, 4
SSDs provided up to two times the performance of HDDs in the ABAQUS
tests
Sun used the following configuration for the ABAQUS test comparisons.
Hardware configuration
• Sun Fire X4450 server
• Four 2.93 GHz quad-core Intel Xeon Processor X7350 CPUs
• Four 15,000 RPM 500 GB SAS drives
• Three 32 GB SSDs
The system was set up to boot off of one of the hard disk drives. The base-line hard-
disk based file system was set to stripe across three SAS HDDs. For comparative
purposes, the SSD-based file system was configured across three SSDs.
Software configuration
• 64-bit SUSE Linux Enterprise Server SLES 10 SP 1
• ABAQUS V6.8-1 Standard Module
• ABAQUS 6.7 Standard Benchmark Test Suite
Sun Microsystems, Inc.9 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
2 For more information on the Sun Fire x2270 server, please see http://www.sun.com/servers/x64/x2270/
The NASTRAN benchmark applicationNASTRAN is an FEA program that was originally developed for NASA (National
Aeronautics and Space Administration) in the late 1960s under United States
government funding for the Aerospace industry. NASTRAN is widely used throughout
the world in the aerospace, automotive, and maritime industries.
The MSC/ NASTRAN test suite was used to compare the performance of a Sun Fire
server using either HDDs or SSDs. Runs were made on a Sun Fire x2270 server2 using
from one to eight cores per job with the MSC/NASTRAN Vendor 2008 benchmark test
suite. (Please see “NASTRAN benchmark test cases” on page 27 for details.) In some
cases only one core was used since some test cases don't scale well beyond this
point. A few scaled well up to four cores, and the rest scaled well up to the eight
cores that were used for this report. The test cases for the MSC/NASTRAN module
have a substantial I/O component where from 15% to 25% of the total run times
could be associated with I/O activity (primarily scratch files).
The Sun Fire x2270 server equipped with two 2.93 GHz quad-core Intel Xeon
Processor X5570 CPUs demonstrated a substantial performance improvement using
Sun SSDs as compared to HDDs. The performance increased in concert with the
system load (increased number of active cores).
Charting these MSC/NASTRAN Vendor_2008 test suite runs to the particular test
as shown in Table 2, one can see greater than two times the overall speed and
productivity of the xxocmd2 eight-core as well as the xlotdf1 eight-core MCS/
NASTRAN benchmark tests. So as the number of cores increased, the overall clock
time of the runs decreased, for an overall majority of the test suite. The xxocmd2
and xlotdf1 eight-core runs each increased performance by more then 54%.
Sun Microsystems, Inc.10 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
Table 2. MSC/NASTRAN Test Suite: HDD vs SSD
Test--number of
cores
Sun Fire
x2270 server
Time (sec)
x2270 HDD
Sun Fire
x2270 server
Time (sec)
x2270 SSD
Time Ratio
HDD:SSD
Improvment Number
Cores
vlosst1-1 127 126 1.007936508 0.79% 1
xxocmd2-1 895 884 1.012443439 1.23% 1
xxocmd2-2 614 583 1.053173242 5.05% 2
xxocmd2-4 631 404 1.561881188 35.97% 4
xxocmd2-8 1554 711 2.185654008 54.25% 8
xlotdf1-1 2000 1939 1.031459515 3.05% 1
xlotdf1-2 1240 1189 1.042893188 4.11% 2
xlotdf1-4 833 751 1.10918775 9.84% 4
xlotdf1-8 1562 712 2.193820225 54.42% 8
sol400_1-1 2479 2402 1.032056619 3.11% 1
sol400_S-1 2450 2262 1.08311229 7.67% 1
getrag-1 843 817 1.031823745 3.08% 1
The testing shows a significant gain in productivity for MCS/NASTRAN when using
SSDs. As seen in Figure 4, the MSC/NASTRAN test suite demonstrates significant
improvement in clock time in nearly all cases, with a gain of nearly two times in the
xxocmd2 and xlotdf1 tests with eight cores.
Sun Microsystems, Inc.11 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
Figure 4.
0
500
1000
1500
2000
2500
1 2 3 4 5 6 7 8 9 10 11 12
Tim
e (s
ec)
Test - #cores
HDD
SSD
vlosst
1 - 1
xxocm
d2 - 1
xxocm
d2 - 2
xxocm
d2 - 4
xxocm
d2 - 8
xlotdf1 - 1
xlotdf1 - 2
xlotdf1 - 4
Xlotdf1 - 8
Sol400_1 - 1
Sol400_S -
1
getrag - 1
SSDs improved performance by more than 54% in the xxocmd2 and xlotdf1 MSC/NASTRAN test
Sun used the following configuration for the NASTRAN test comparisons:
Hardware configuration
• Sun Fire x2270 server
• Two 2.93 GHz quad-core Intel Xeon Processor X5570 CPUs
• 24 GB memory
• Three 7200 RPM SATA HDD
• Two 32 GB SSD
The system was set up to boot off of one of the hard disk drives. The base-line
hard-disk based file system was set to stripe across two SATA HDDs. For comparative
purposes, the SSD-based file system was configured across both SSDs.
Software configuration
• 64-bit SUSE Linux Enterprise Server SLES 10 SP 1
• MSC/NASTRAN MD 2008
• MSC/NASTRAN Vendor_2008 Benchmark Test Suite
Sun Microsystems, Inc.12 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
The ANSYS benchmark applicationANSYS is a general-purpose FEA modeling package used widely in industry. The
ANSYS BMD Test Suite was used to acquire this data. (Please see “ANSYS 12.0 (prel. 7)
with ANSYS 11.0 distributed benchmarks” on page 29 for details.) This test suite was
used to compare the performance of a Sun Fire server equipped with HDDs and SSDs.
Runs were made on a Sun Fire x2270 server using eight cores per job. The test cases
have a substantial I/O component where 15% to 20% of the total run times are
associated with I/O activity (primarily scratch files).
The Sun Fire x2270 server equipped with two 2.93 GHz quad-core Intel Xeon
Processor X5570 CPUs demonstrated a substantial performance improvement using
Sun SSDs as compared to using the traditional HDDs. One of the most I/O intensive
cases in the ANSYS BMD test suite is the bmd-4 case. This test case in particular
showed a significant increase in overall performance and productivity. The same test
running with HDDs took 2.78 times longer to complete than when the system was
equipped with SSDs.
Table 3 illustrates the overall comparisons for the ANSYS BMD test suite runs:
Table 3. HDD-based system required as much as 2.78 times longer as SSD on the ANSYS BMD test suite
Eight-core
BM test
Sun Fire
x2270 server
Time (sec)
x2270 HDD
Sun Fire
x2270 server
Time (sec)
x2270 SSD
Time Ratio
HDD:SSD
Improvement
bmd-1 39 26 1.5 33.33%
bmd-2 117 84 1.392857143 28.21%
bmd-3 68 66 1.03030303 2.94%
bmd-4 703 253 2.778656126 64.01%
bmd-5 298 285 1.045614035 4.36%
bmd-6 297 292 1.017123288 1.68%
bmd-7 293 212 1.382075472 27.65%
As shown in Figure 5, the ANSYS BMD test suite bmd-4 runs using SSDs yield 2.78
times the overall speed and productivity of the same test using HDDs. The bmd-4
eight-core run improved by 64.01% simply by using SSDs.
Sun Microsystems, Inc.13 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
Figure 5.
0
100
200
300
400
500
600
700
800
1 2 3 4 5 6 7
Tim
e (s
ec)
Test
HDD
SSD
bmd-1 bmd-2 bmd-3 bmd-4 bmd-5 bmd-6 bmd-7
The bmd-4 test eight-core run improves performance by 64.01%
The testing shows a substantial boost in productivity for ANSYS when using SSDs.
Sun used the following configuration for the ANSYS test comparisons described in
this report:
Hardware configuration
• Sun Fire x2270 server
• Two 2.93 GHz quad-core Intel Xeon Processor X5570 CPUs
• 24 GB memory
• Two 32 GB SSDs
• Three 7200 rpm SATA 500 GB HDDs
The system was set up to boot from one of the hard disk drives. The base-line hard-
disk based file system was set to stripe across two SATA HDDs. For comparative
purposes, the SSD-based file system was configured across two SSDs.
Software configuration
• 64-bit SUSE Linux Enterprise Server SLES 10 SP 2
• ANSYS V 12.0 Prerelease 7
• ANSYS 11 Distributed BMD Benchmark Test Suite
Sun Microsystems, Inc.14 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
3 HPC User Forum Survey, 2007 HPC Storage and Data Management: User/Vendor Perspectives and Survey Resuls
Summary for single system application performanceThese tests have demonstrated that the use of SSDs can lead to overall improvement
in performance in testing using HPC MCAE applications when compared to the
same applications run on HDDs. This improvement was seen using both SAS and
SATA-based configurations. These tests have demonstrated that SSDs can improve
performance markedly in I/O bound applications. It is also important to note that
the use of SSDs can result in a reduction of overall power consumption. The systems
run cooler and have less of an impact on the environment.
This testing demonstrated that the greatest reduction in wall-clock time, and
improvement in productivity, is associated with benchmark applications that
have the most significant I/O component. In cases where the I/O load is less, as
expected, the performance improvement is more limited. In the case of the ANSYS
bmd-5 benchmark, for example, if sufficient memory is available, the solver can
run in memory. In this case, no I/O is required at all, and the improvement in I/O
bandwidth has little or no effect on the performance of the benchmark application.
Thus, it is important to consider the I/O requirements of a particular application
when considering the use of SSD to improve performance.
SSD usage with the Lustre™ parallel file systemThe Lustre parallel file system is an open source, shared file system designed to
address the I/O needs of the largest and most demanding compute clusters. The
Lustre parallel file system is best known for powering the largest HPC clusters in the
world, with tens of thousands of client systems, petabytes of storage, and hundreds
of gigabytes per second of I/O throughput. A number of HPC sites use the Lustre file
system as a site-wide global file system, servicing clusters on an exceptional scale.
The Lustre file system is used by over 40% of Top 100 Supercomputers as ranked by
top500.org on the November 2008 listing. Additionally, IDC lists the Lustre file system
as the file system with the largest market share in HPC.3
With the mass adoption of clusters and explosive growth of data storage needs,
I/O bandwidth challenges are becoming common in a variety of public and private
sector environments. The Lustre file system is a natural fit for these situations where
traditional shared file systems, such as NFS, do not scale to the required aggregate
throughput requirements. Sectors struggling with this challenge can include oil and
gas, manufacturing, government, and digital content creation (DCC).
Lustre file system designThe Lustre file system (Figure 6) is a software-only architecture that allows a number
of different hardware implementations. The main components of the Lustre file
system architecture are Lustre file system clients (Lustre clients), Metadata Servers
(MDS), and Object Storage Servers (OSS). Lustre clients are typically compute nodes
Sun Microsystems, Inc.15 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
in HPC clusters. These nodes run Lustre client software, and access the Lustre file
system via InfiniBand, Gigabit Ethernet, or 10 Gigabit Ethernet connections. The
Lustre file system client software presents a native POSIX file interface to the client
nodes on which it runs. The Lustre file system is then mounted like any other file
system. Metadata Servers and Object Storage Servers implement the file system and
communicate with the Lustre clients.
Figure 6.
MetadataServers(MDS)
(active) (standby)
Object Storage Servers (OSS)
Storage Arrays (Direct Connect)
Enterprise Storage Arrays & SAN Fabrics
Commodity Storage
Ethernet
Multiple networkssupported
simultaneously
Clients
File SystemFail-over
InfiniBand
The Lustre file system
The Lustre file system uses an object-based storage model, and provides several
abstractions designed to improve both performance and scalability. At the file
system level, Lustre file system technology treats files as objects that are located
through metadata servers. Metadata servers support all file system name space
operations, such as file lookups, file creation, and file and directory attribute
manipulation. File data is stored in objects on the OSSs. The MDS directs actual
file I/O requests from Lustre file system clients to OSSs, which ultimately manage
Sun Microsystems, Inc.16 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
4 http://www.sun.com/software/products/hpcsoftware/index.xml5 http://www.sun.com/servers/blades/x6250/6 http://www.sun.com/servers/blades/6000/
the storage that is physically located on underlying storage devices. Once the MDS
identifies the storage location of a file, all subsequent file I/O is performed between
the client and the OSS.
This design divides file system updates into two distinct types of operations: file
system metadata updates on the MDS, and actual file data updates on the OSS.
Separating file system metadata operations from actual file data operations not only
improves immediate performance, but also improves long-term aspects of the file
system such as recoverability and availability.
The Lustre file system implementation supports InfiniBand or Gigabit Ethernet
interconnects, redundant metadata servers, and a choice of commodity storage for
use on the Object Storage Servers. This can include:
• Simple disk storage devices (colloquially, “just a bunch of disks”, or “JBOD”)
• High availability direct storage
• Enterprise SANs
Since the Lustre file system is so flexible, it can be used in place of a shared SAN
for enterprise storage requirements. However, the Lustre file system is also well
suited to use in a traditional SAN environment as well. Lustre file system clusters
are composed of rack mount or blade server clients, metadata servers, and object
storage servers. The Lustre file system runs on Sun’s Open Network Systems
Architecture.
Note: For more information on the Lustre file system, see http://wiki.lustre.org/ and http://www.sun.com/software/products/lustre/.
With the previously documented success on single system runs with SSDs, Sun has
begun to explore using SSDs with the Lustre file system. Testing within Sun has
been performed with a cluster deployed through the use of Sun HPC Software, Linux
Edition4. This software fully integrates the Lustre file system as an open software
component. It also includes OFED (Open Fabrics Enterprise Distribution) software for
Mellanox InfiniBand support.
The test cluster configuration included:
• One Sun Fire x2250 server configured as a Lustre file system client
• One Sun Fire X2250 server configured as an MDS
• Two Sun Blade™ x6250 server modules5 configured with HDDs and SSDs as OSSs,
Sun Blade 6000 Modular System enclosure6
• One Dual Data Rate (DDR) InfiniBand Network
Sun Microsystems, Inc.17 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
7 http://jperfmeter.sourceforge.net/
After the systems were provisioned with the Sun HPC Software, Linux Edition a Lustre
file system was created using the commands below.
The following commands were executed to configure the MDS server.
mkfs.lustre --fsname=testfs --mgs --mdt /dev/sdbmkdir mdtmount -t lustre /dev/sdb /mdt
The first Sun Blade X6250 server module acting as an OSS was configured with file
systems based on both a HDD and an SSD.
mkfs.lustre --fsname=testfs --ost --mgsnode=v6i@o2ib /dev/sdamkfs.lustre --fsname=testfs --ost --mgsnode=v6i@o2ib /dev/sdcmkdir ostsdahddmkdir ostsdcssdmount -t lustre /dev/sda /ostsdahddmount -t lustre /dev/sdc /ostsdcssd
The second Sun Blade X6250 server module acting as an OSS was configured with a
single SSD-based file system.
mkfs.lustre --fsname=testfs --ost --mgsnode=v6i@o2ib /dev/sdcmkdir ostsdcssdmount -t lustre /dev/sdc /ostsdcssd
The Lustre file system client was then configured to access the various HDD-based
and SSD-based file systems for testing.
mkdir lustrefsmount -t lustre v6i@o2ib:/testfs /lustrefsmkdir /lustrefs/st1hddmkdir /lustrefs/st1ssdmkdir /lustrefs/st2ssdlfs setstripe /lustrefs/st1hdd -c 1 -s 1m -i 1lfs setstripe /lustrefs/st1ssd -c 1 -s 1m -i 2lfs setstripe /lustrefs/st2ssd -c 2 -s 1m -i 2
Note: The Lustre file system command lfs setstripe was used on specific directories (st1hdd, st1ssd, st2ssd) to direct I/O to specific HDDs and SSDs for data contained in
this report.
In addition, the lfs getstripe Lustre file system command was used to review
that proper striping was in force as well as specific object-storage targets (OSTs)
were assigned that were needed to support the specific tests that were run in this
report as described. The Java™ Performance statistics monitor (JPerfmeter7) was also
incorporated to see which OSS/OST was being used.
Sun Microsystems, Inc.18 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
8 http://www.iozone.org/
IOZone file system testingThe IOZone file system benchmark tool8 is used to perform broad--based performance
testing of file systems, using a synthetic workload with a wide variety of file system
operations. IOZone is an independent, portable benchmark that is used through the
industry.
Runs were first made with the HDD-based OSS and the IOZone benchmark in order to
establish baseline performance. Similar runs were then made using the SSD-based
configuration, again recording performance using IOZone.
From the Lustre file system client several IOZone commands were used to gather
data for these tests. The following IOZone command was used to direct traffic to the
HDD-based OSS on the first Sun Blade X6250 server module.
iozone -r 1024k -s 2G -i 0 -i 1 -f /lustrefs/st1hdd/iozone2ghdd -Rb /lustrefs/st1hdd/iozone2ghdd.xls -+m /lustrefs/scripts/iozone/iozone3_311/src/current/client_list
The following IOZone command was used to direct traffic to the SSD-based OSS on
the first Sun Blade X6250 server module.
iozone -r 1024k -s 2G -i 0 -i 1 -f /lustrefs/st1ssd/iozone2gssd -Rb /lustrefs/st1ssd/iozone2gssd.xls -+m /lustrefs/scripts/iozone/iozone3_311/src/current/client_list
The following IOZone command was used to direct traffic to the two separate SSD-
based OSSs on each of the Sun Blade X6250 server modules. This testing was done
to verify that scaling would occur with the Lustre file system and multiple SSD-based
OSSs.
iozone -r 1024k -s 2G -i 0 -i 1 -f /lustrefs/st2ssd/iozone2g-2-ssd -Rb /lustrefs/st2ssd/iozone2g-2-ssd.xls -+m /tmp/lustrefs/scripts/iozone/iozone3_311/src/current/client_list
Note: These are single command lines, reformatted to fit the page.
Sun Microsystems, Inc.19 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
Figure 7 shows data write performance with the results of:
• A single HDD-based OSS as the baseline
• A single SSD-based OSSs for initial comparison
• Two SSD-based OSSs to verify scaling
Figure 7. Data write performance was greater with SSDs, and scaled with multiple SSD-based OSSs
A Lustre file system using SSDs shows significant advantages over a similar Lustre file
system using the baseline HDD configuration:
• Using a Lustre file system configuration with a single OSS using SSD, runs required
only 77.5% of the time required using the baseline HDD configuration.
• I/O bandwidth was 1.37 greater using the Lustre file system in a single SSD OSS
configuration, compared to the baseline HDD OSS.
• Using a Lustre file system configuration with two OSSs using SSD, run time was
reduced to 41.65% of the time required using the baseline HDD configuration.
• I/O bandwidth was 2.39 times greater using the Lustre file system with two OSS/
SSD devices, compared to the baseline HDD OSS.
These results demonstrate that SSDs provide improved performance used in an
OST for the Lustre file system. Two OSS/SSD devices show further improvement,
demonstrating that OSS/SSD performance scales with the number of OSSs
Sun Microsystems, Inc.20 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
9 http://wikis.sun.com/display/BluePrints/Solving+the+HPC+IO+Bottleneck+-+Sun+Lustre+Storage+System
Sun used the following configuration for the OSS server to run the IOZone test
comparisons described in this report:
Hardware configuration
• Sun Blade X6250 server module
• Two 3.00 GHz quad-core Intel Xeon Processor E5450 CPUs
• 16 GB memory
• One SSD
• One 10000 RPM SAS HDD
Software configuration
• Red Hat Enterprise Linux 5.1 CentOS 5.1 x86_64
• 2.6.18-53.1.14.el5_lustre.1.6.5smp
• The IOZone file system benchmarking tool
Summary for SSD usage with the Lustre parallel file systemThis report has shown that the use of SSD-based OSSs can drive I/O faster than
traditional HDD-based OSSs. Testing showed, further, that the Lustre file system
can scale with the use of multiple SSD-based OSSs. Not only can I/O bandwidth be
increased with the use of the Lustre file system and SSDs but it is anticipated that
run times of other applications using the Lustre file system equipped with SSDs can
also be reduced .
Future directionsNew technology included in the Lustre file system version 1.8 allows pools of
storage to be configured based on technology and performance, and then allocated
according to the needs of specific jobs. So, for example, an elite pool of extremely
fast SSD storage could be defined along with pools of slower, but higher capacity,
HDD storage. Other pools might be defined to use local devices, SAN devices,
or networked file systems. The Lustre file system then allows these pools to be
allocated as needed to specific jobs in order to optimize performance based upon
service level objectives.
Performance studies of a production Lustre file system have been performed at the
Texas Advanced Computing Center (TACC), using the scaling capabilities of the Lustre
file system to obtain higher performance, therefore reducing the I/O bottleneck.
(This work is described in the Sun BluePrint Solving the HPC I/O Bottleneck: Sun
Lustre Storage System9.)
Future work will explore the use of SSDs integrated with new versions of the Lustre
file system, Quad Data Rate (QDR) InfiniBand, and Sun’s new servers and blades.
Sun Microsystems, Inc.21 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
ConclusionUse of SSDs with Sun servers and blades has demonstrated significant performance
improvements in single-system runs of FEA HPC application benchmarks, and
through the use of the Lustre parallel file system. There is significant promise that
other applications with similar data throughput needs and workloads will also
obtain increased bandwidth as well as a reduction in run times.
Sun Microsystems, Inc.22 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
Appendix: Benchmark descriptions and parametersThe results reported in this article make use of a collection of benchmark numerical
applications. Each benchmark suite makes particular requirements for data that
should be made available so the benchmarks can be evaluated fairly. In this
Appendix, we note the required details for each of the benchmarks used.
ABAQUS standard benchmark test casesThe problems described below provide an estimate of the performance that can
be expected when running ABAQUS/Standard on different computers. The jobs are
representative of typical ABAQUS/Standard applications including linear statics,
nonlinear statics, and natural frequency extraction.
• S1: Plate with gravity load
This benchmark is a linear static analysis of a plate with gravity loading. The plate
is meshed with second-order shell elements of type S8R5 and uses a linear elastic
material model. Edges of the plate are fixed. There is no contact.
– Input file name: s1.inp
– Increments: 1
– Iterations: 1
– Degrees of freedom: 1,085,406
– Floating point operations: 1.89E+011
– Minimum memory requirement: 587 MB
– Memory to minimize I/O: 2 GB
– Disk space requirement: 2 GB
• S2: Flywheel with centrifugal load
This benchmark is a mildly nonlinear static analysis of a flywheel with centrifugal
loading. The flywheel is meshed using first-order hexahedral elements of type
C3D8R and uses an isotropic hardening Mises plasticity material model. There
is no contact. The nonlinearity in this problem arises from localized yielding in
the vicinity of the bolt holes. Two versions of this benchmark are provided. Both
versions are identical except that one uses the direct sparse solver and the other
uses the iterative solver.
• S2a: Direct solver version
– Input file name: s2a.inp
– Increments: 6
– Iterations: 12
– Degree of freedom: 474,744
– Floating point operations: 1.86E+012
– Minimum memory requirement: 733 MB
– Memory to minimize I/O: 849 MB
– Disk space requirement: 4.55 GB
Sun Microsystems, Inc.23 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
• S2b: Iterative solver version
– Input file name: s2b.inp
– Increments: 6
– Iterations: 11
– Degrees of freedom: 474,744
– Floating point operations: 8.34E+010
– Minimum memory requirement: 2.8 GB
– Memory to minimize I/O: NA
– Disk space requirement: 387 MB
• S3: Impeller frequencies
This benchmark extracts the natural frequencies and mode shapes of a turbine
impeller. The impeller is meshed with second-order tetrahedral elements of type
C3D10 and uses a linear elastic material model. Frequencies in the range from 100
Hz. to 20,000 Hz. are requested.
Three versions of this benchmark are provided: a 360,000 DOF version that
uses the Lanczos eigensolver, a 1,100,000 DOF version that uses the Lanczos
eigensolver, and a 1,100,000 DOF version that uses the AMS eigensolver.
• S3a: 360,000 DOF Lanczos eigensolver version
– Input file name: s3a.inp
– Degrees of freedom: 362,178
– Floating point operations: 3.42E+11
– Minimum memory requirement: 384 MB
– Memory to minimize I/O: 953 MB
– Disk space requirement: 4.0 GB
• S3b: 1,100,000 DOF Lanczos eigensolver version
– Input file name: s3b.inp
– Degrees of freedom: 1,112,703
– Floating point operations: 3.03E+12
– Minimum memory requirement: 1.33 GB
– Memory to minimize I/O: 3.04 GB
– Disk space requirement: 23.36 GB
• S3c: 1,100,000 DOF AMS eigensolver version
– Input file name: s3c.inp
– Degrees of freedom: 1,112,703
– Floating point operations: 3.03E+12
– Minimum memory requirement: 1.33 GB
– Memory to minimize I/O: 3.04 GB
– Disk space requirement: 19.3 GB
Sun Microsystems, Inc.24 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
• S4: Cylinder head bolt-up
This benchmark is a mildly nonlinear static analysis that simulates bolting a
cylinder head onto an engine block. The cylinder head and engine block are
meshed with tetrahedral elements of types C3D4 or C3D10M, the bolts are meshed
using hexahedral elements of type C3D8I, and the gasket is meshed with special-
purpose gasket elements of type GK3D8. Linear elastic material behavior is used
for the block, head, and bolts while a nonlinear pressure-overclosure relationship
with plasticity is used to model the gasket. Contact is defined between the bolts
and head, the gasket and head, and the gasket and block. The nonlinearity in this
problem arises both from changes in the contact conditions and yielding of the
gasket material as the bolts are tightened.
Three versions of this benchmark are provided: a 700,000 DOF version that is
suitable for use with the direct sparse solver on 32-bit systems, a 5,000,000 DOF
version that is suitable for use with the direct sparse solver on 64-bit systems, and
a 5,000,000 DOF version that is suitable for use with the iterative solver on 64-bit
systems.
• S4a: 700,000 DOF direct solver version
– Input file name: s4a.inp
– Increments: 1
– Iterations: 5
– Degrees of freedom: 720,059
– Floating point operations: 5.77E+11
– Minimum memory requirement: 895 MB
– Memory to minimize I/O: 3 GB
– Disk space requirement: 3 GB
• S4b: 5,000,000 DOF direct solver version
– Input file name: s4b.inp
– Increments: 1
– Iterations: 5
– Degrees of freedom: 5,236,958
– Floating point operations: 1.14E+13
– Minimum memory requirement: 4 GB
– Memory to minimize I/O: 20 GB
– Disk space requirement: 23 GB
• S4c: 5,000,000 DOF iterative solver version
– Input file name: s4c.inp
– Increments: 1
– Iterations: 3
– Degrees of freedom: 5,248,154
– Floating point operations: 3.74E+11
Sun Microsystems, Inc.25 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
– Minimum memory requirement: 16 GB
– Memory to minimize I/O: NA
– Disk space requirement: 3.3 GB
• S5: Stent expansion
This benchmark is a strongly nonlinear static analysis that simulates the
expansion of a medical stent device. The stent is meshed with hexahedral
elements of type C3D8 and uses a linear elastic material model. The expansion
tool is modeled using surface elements of type SFM3DR. Contact is defined
between the stent and expansion tool. Radial displacements are applied to the
expansion tool which in turn cause the stent to expand. The nonlinearity in this
problem arises from large displacements and sliding contact.
– Input file name: s5.inp
– Increments: 21
– Iterations: 91
– Degrees of freedom: 181,692
– Floating point operations: 1.80E+009
– Minimum memory requirement: NA
– Memory to minimize I/O: NA
– Disk space requirement: NA
Note: Abaqus, Inc. would like to acknowledge Nitinol Devices and Components for providing the original finite element model of the stent. The stent model used in this benchmark is not representative of current stent designs.
• S6: Tire footprint
This benchmark is a strongly nonlinear static analysis that determines the
footprint of an automobile tire. The tire is meshed with hexahedral elements of
type C3D8, C3D6H, and C3D8H. Linear elastic and hyperelastic material models
are used. Belts inside the tire are modeled using rebar layers and embedded
elements. The rim and road surface are modeled as rigid bodies. Contact is
defined between the tire and wheel and the tire and road surface. The analysis
sequence consists of three steps. During the first step the tire is mounted to the
wheel, during the second step the tire is inflated, and then during the third step a
vertical load is applied to the wheel. The nonlinearity in the problem arises from
large displacements, sliding contact, and hyperelastic material behavior.
– Input file name: s6.inp
– Increments: 41
– Iterations: 177
– Degrees of freedom: 729,264
– Floating point operations: NA
Sun Microsystems, Inc.26 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
– Minimum memory requirement: 397 MB
– Memory to minimize I/O: 940 MB
– Disk space requirement: NA
Hardware configuration
• Sun Fire X4450 server
• Four 2.93 GHz quad-core Intel Xeon X7350 Processor CPUs
• Four 15,000 RPM 500 GB SAS drives
• Three 32 GB SSDs
The system was set up to boot off of one of the hard disk drives. The base-line hard-
disk based file system was set to stripe across three SAS HDDs. For comparative
purposes, the SSD-based file system was configured across three SSDs.
Software configuration
• 64-bit SUSE Linux Enterprise Server SLES 10 SP 1
• ABAQUS V6.8-1 Standard Module
• ABAQUS 6.7 Standard Benchmark Test Suite
Sun Microsystems, Inc.27 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
NASTRAN benchmark test casesThe problems described below are representative of typical MSC/Nastran
applications including both SMP and DMP runs involving linear statics, nonlinear
statics, and natural frequency extraction.
• vl0sst1
– No. Degrees Of Freedom: 410,889
Run time sensitive to memory allocated to job:
– 2:04:36 elapsed w/ mem=37171200
– 4:35:26 elapsed w/ mem=160mb sys1=32769
– 5:20:12 elapsed w/ mem=80mb sys1=32769
– 1:11:58 elapsed w/ mem=1600mb bpool=40000
(This job does extensive post solution processing of GPSTRESS I/O. )
– Solver: SOL 101
– Memory Usage: 7.3 MB
– Maximum Disk Usage: 4.33 GB
• xx0cmd2
– No. Degrees Of Freedom: 1,315,562
– Solver: SOL 103
– Normal Modes With ACMS - DOMAINSOLVER ACMS (Automated Component
Modal Synthesis)
– Memory Usage: 1800 MB
– Maximum Disk Usage: 14.422 GB
• xl0tdf1
– No. Degrees Of Freedom: 529,257
– Solver: SOL 108 Fluid/Solid Interaction
– Car Cabin Noise - FULL VEHICLE SYSTEM MODEL
– Eigenvalue extraction - Direct Frequency Response
– Memory Usage: 520 MB
– Maximum Disk Usage: 5.836 GB
• xl0imf1
– No. Degrees Of Freedom: 468,233
– Fluid/Solid Interaction
– Frequency Response
– Memory Usage: 503 MB
– Maximum Disk Usage: 10.531 GB
Sun Microsystems, Inc.28 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
• md0mdf1
– No. Degrees Of Freedom: 42,066
– This model is for Exterior Acoustics
– Modal Frequency Response Analysis With UMP Pack
– Fluid/Solid Interaction
– Memory Usage: 1 GB
– Maximum Disk Usage: 414.000 MB
• 400_1 & 400_S
– No. Degrees Of Freedom: 437,340
– Solver: 400 (MARC module)
– Nonlinear Static Analysis
– Memory Usage: 1.63 GB
– Maximum Disk Usage: 3.372 GB
(S Model Sets Aside 3 GB Physical Memory For I/O Buffering)
• getrag (Contact Model)
– No. Degrees Of Freedom: 2,450,320
– PCGLSS 6.0: Linear Equations Solver
– Solver: 101
– Memory Usage: 8.0 GB
– Maximum Disk Usage: 17.847 GB
– Total I/O: 139 GB
Hardware configuration
• Sun Fire x2270 server
• Two 2.93 GHz quad-core Intel Xeon Processor X5570 CPUs
• 24 GB memory
• Three 7200 RPM SATA 500 GB HDDs
• Two 32 GB SSDs
The system was set up to boot from one of the hard disk drives. The base-line hard-
disk based file system was set to stripe across two SATA HDDs. For comparative
purposes, the SSD-based file system was configured across both SSDs.
Software configuration
• 64-bit SUSE Linux Enterprise Server SLES 10 SP 1
• MSC/NASTRAN MD 2008
• MSC/NASTRAN Vendor_2008 Benchmark Test Suite
Sun Microsystems, Inc.29 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
ANSYS 12.0 (prel. 7) with ANSYS 11.0 distributed benchmarks
• bmd-1
– Dsparse solver, 400K DOF
– Static analysis
– Medium sized job, should run in-core on all systems
• bmd-2
– 1M DOF iterative solver job.
– Shows good scaling due to simple preconditioner
• bmd-3
– 2M DOF Static analysis
– Shows good parallel performance for iterative solver
– Uses pcg iterative solver
– Uses msave,on feature, cache friendly
• bmd-4
– Larger dsparse solver job
– 3M DOF, tricky job for dsparse when memory is limited
– Shows I/O as well as CPU performance
– Good to show benefit of large memory
• bmd-5
– 5.8M DOF large pcg solver job
– Good parallel performance for iterative solver on a larger job
– Cache friendly msave,on elements
• bmd-6
– 1M DOF lanpcg: Uses assembled matrix with PCG preconditioner
– New iterative modal based analysis solver chosen to maximize speedups
• bmd-7
– 5M DOF static analysis, uses solid45 elements
– Best test of memory bandwidth performance, which are NOT msave,on
elements
– Lower mflop rate is expected because of sparse matrix/vector kernel
Sun Microsystems, Inc.30 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
Hardware configuration
• Sun Fire x2270 server
• Two 2.93 GHz quad-core Intel Xeon Processor X5570 CPUs
• 24 GB memory
• Two 32 GB SSDs
• Three 7200 rpm SATA 500 GB HDDs
The system was set up to boot from one of the hard disk drives. The base-line hard-
disk based file system was set to stripe across two SATA HDDs. For comparative
purposes, the SSD-based file system was configured across three SSDs.
Software configuration
• 64-bit SUSE Linux Enterprise Server SLES 10 SP 2
• ANSYS V 12.0 Prerelease 7
• ANSYS 11 Distributed BMD Benchmark Test Suite
About the authorsLarry McIntosh is a Principal Systems Engineer at Sun Microsystems and works
within Sun’s Systems Engineering Solutions Group. He is responsible for designing
and implementing high performance computing technologies at Sun’s largest
customers. Larry has 35 years of experience in the computer, communications,
and storage industries and has been a software developer and consultant in the
commercial, government, education and research sectors as well as a computer
science college professor. Larry’s recent work has included the deployment of the
Ranger system servicing the National Science Foundation and Researchers at the
Texas Advanced Computer Center (TACC) in Austin, Texas.
Michael Burke obtained his Ph.D. from Stanford University. Since then he has spent
over 35 years in the development and application of MCAE software. He was the
principal developer of the MARC code now owned by MSC/Nastran. Following the SS
Challenger disaster he developed FANTASTIC (Failure Analysis Thermal and Structural
Integrated Code) for NASA and its suppliers/contractors for the analysis of rocket
(nozzles) More recently he has been involved with the benchmarking of state of the
art HPC platforms using the more prominent commercial ISV MCAE/CFD/CRASH and
other scientific applications He has performed this benchmarking for Fujitsu and
Hewlett Packard, and is currently in the Strategic Applications Engineering group at
Sun Microsystems.
Sun Microsystems, Inc.31 Solid State Drives in HPC: Reducing the I/O Bottleneck
.
References
Web Sites
Sun Fire x4450 server http://www.sun.com/servers/x64/x4450/
Sun Fire x2270 server http://www.sun.com/servers/x64/x2270/
Sun HPC Software, Linux Edition http://www.sun.com/software/products/
hpcsoftware/index.xml
Sun Blade x6250 server module http://www.sun.com/servers/blades/x6250/
Sun Blade 6000 Modular System chassis http://www.sun.com/servers/blades/6000/
JPerfMeter http://jperfmeter.sourceforge.net/
IOZone Benchmark http://www.iozone.org/
Sun BluePrints Articles
Solving the HPC I/O Bottleneck: Sun
Lustre Storage System
http://wikis.sun.com/display/BluePrints/
Solving+the+HPC+IO+Bottleneck+-
+Sun+Lustre+Storage+System
Ordering Sun DocumentsThe SunDocsSM program provides more than 250 manuals from Sun Microsystems,
Inc. If you live in the United States, Canada, Europe, or Japan, you can purchase
documentation sets or individual manuals through this program.
Accessing Sun Documentation OnlineThe docs.sun.com Web site enables you to access Sun technical documentation
online. You can browse the docs.sun.com archive or search for a specific book title or
subject. The URL is http://docs.sun.com
To reference Sun BluePrints Online articles, visit the Sun BluePrints Online Web site
at: http://www.sun.com/blueprints/online.html
Sun Microsystems, Inc.
Sun Microsystems, Inc. 4150 Network Circle, Santa Clara, CA 95054 USA Phone 1-650-960-1300 or 1-800-555-9SUN (9786) Web sun.com
Solid State Drives in HPC: Reducing the I/O Bottleneck
© 2009 Sun Microsystems, Inc. All rights reserved. Sun, Sun Microsystems, the Sun logo, Java, Sun Blade, and Sun Fire are trademarks or registered trademarks of Sun Microsystems, Inc. or its subsidiaries in the United States and other countries. Intel Xeon is a trademark or registered trademark of Intel Corporation or its subsidiaries in the United States and other countries. Information subject to change without notice. Printed in USA 06/2009
Top Related