Solution Brief: Life Sciences - HPC Advisory Council...Solution Brief: Higher Education and Research...
Transcript of Solution Brief: Life Sciences - HPC Advisory Council...Solution Brief: Higher Education and Research...
Solution Brief:
Life SciencesOpening the pipe to faster research, discovery, computation and resource sharing
Abstract
Advances in Information Technology (IT) are significantly improving the speed at which organizations
dedicated to improving the lives of many through research and development conduct business today. In
order to dramatically improve the efficiency of their computing infrastructure, life sciences organizations
and educational institutions helping with scientific research, computational chemistry, systems biology, and
chemical mixing and analysis are turning to Voltaire’s InfiniBand-based solutions.
The Life Sciences Challenge
There are many focus areas for research and development in the life sciences. Life science IT managers are
challenged to provide the right solutions for computational chemistry, systems biology, and chemical mixing
and analysis. These areas share several challenges:
1. Conducting more simulations per day
2. Alleviating storage bottlenecks associated with exponential data growth
3. Implementing cost effective solutions
4. Allocating compute and storage resources dynamically to meet researchers’ needs
Complex simulations can take days or weeks to run. When simulations take longer, scientific discoveries are
delayed, slowing commercialization and increasing competitive threats. With faster simulations, more complex
models can be analyzed, additional assumptions can be tested and further modifications become possible
leading to more efficient development cycles.
Accelerating simulations has clear benefits. The key to accelerating life science application performance is
to select high performance computing systems that eliminate bottlenecks associated with Inter-Processor
Communications (IPC) and storage connectivity.
Large symmetric multi-processing-machines (SMPs) have been used in the past as an answer for generating
massive compute power in data centers. However, these proprietary, expensive systems have given way to
cluster and grid architectures built from lower-cost commodity elements that offer incredible performance at a
significantly lower cost.
Because of the ready-availability of Ethernet, many of today’s clusters and grids are built with Ethernet as the
interconnect. While Gigabit Ethernet-based clustering is less expensive than SMP-based architectures, it tends
to be very inefficient. For applications that rely on bandwidth or memory sharing, or require large amounts of
data transfer, InfiniBand-based interconnects significantly improve the efficiency of clusters by providing high
bandwidths and low latency without increasing CPU utilization.
Contents
The Life Sciences Challenge . . . . . . . 1
Today’s Solutions . . . . . . . . . . . . . . . . . 2
Computational Chemistry . . . . . . . . 2
Systems Biology . . . . . . . . . . . . . . . 2
Chemical Mixing and Material Analysis . . . . . . . . . . . . . . . . . . . . . . . 2
A Better Way . . . . . . . . . . . . . . . . . . . . . 3
Building High Performance Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . 4
High-performance InfiniBand Switches . . . . . . . . . . . . . . . . . . . . . . 4
Fast Storage Access . . . . . . . . . . . . 7
Visualization Solutions . . . . . . . . . . 8
Putting It All Together . . . . . . . . . . . . . 8
Key Features & Benefits . . . . . . . . . 8
Tested & Certified with Leading Applications . . . . . . . . . . . . . . . . . . . 9
Customer Success Stories . . . . . . . . . 9
Swiss Institute of BioInformatics (SIB) . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Tokyo Institute of Technology (TiTech) . . . . . . . . . . . . . . . . . . . . . . 11
About Voltaire . . . . . . . . . . . . . . . . . . . 12
2
Solution Brief: Higher Education and Research
When clusters are built using high performance servers, storage and interconnects, organizations can
experience drastically faster simulations and modeling while continuing to decrease the cost of providing
computing infrastructure.
Today’s Solutions
Many life sciences organizations employ last generation, less-efficient platforms that use proprietary- or
Ethernet-based server interconnects. This approach does not provide the necessary bandwidth for complex
simulations or data transfer that are so common with life sciences applications today.
In the past, applications were “embarrassingly parallel” and were thought to not require a high speed
inteconnect such as InfiniBand. The latest commerical, open source and university-created applications have
been optimized to take advantage of InfiniBand’s improved performance and data transfer speeds.
Computational ChemistryLife science organizations in the area of computational chemistry conduct drug research and discovery as well
as biochemical analysis and modeling. These areas share five primary challenges:
Growing size of formulas and problems �
Need to accelerate development of new life-saving drugs �
More comprehensive safety and drug interaction identification requirements �
Applications that are not optimized for today’s powerful CPUs �
Severe price sensitivity and cost pressure from various groups �
Simulations and modeling for chemical analysis involves huge formulas and constant number crunching. Until
recently, most solutions used Ethernet as an interconnect. Because of the high CPU overhead related to the
handling of communication requests, servers spend more cycles managing inter-processor communications
than actually solving computational tasks.
Systems BiologyLife science organizations conducting analyses and modeling in many different chemical and molecular areas
share three primary challenges:
Simulations of larger, more complex systems and modeling of cell behaviors require lots of �
bandwidth, low latency, and fast storage access
File size and quantity is growing exponentially leading to I/O bottlenecks �
Grids are used to share data, discoveries and models �
Systems biology involves many researchers that constantly retrieve, change, and replace extremely large
files. This causes a tremendous amount of bottlenecks in the network. Because so much data is shared across
organizations, life sciences organizations use large-scale grids.
Chemical Mixing and Material AnalysisLife science organizations conducting research and discovery in chemical analysis, material, adhesive, flavors
and scent modeling share three primary challenges:
Growing size of formulas and problems cause I/O bottlenecks �
The need to accelerate product development �
Safety and quality issues need to be identified sooner in the process �
3
Similar to computational chemistry, the size of formulas can be quite large and cause huge bottlenecks. If these
bottlenecks can be resolved, more analyses can be done faster, accelerating products to market. Additionally,
issues and safety hazzards can be caught sooner in the product cycle, saving organizations thousands of
dollars.
A Better Way
To improve the speed of life sciences applications, engineers need to optimize the design of high performance
computing systems. IT managers spend a lot of time determining the server CPU to be used in clusters and
grids, but the interconnect that is deployed to transport information between the servers is often ignored. This
is a missed opportunity as less efficient interconnects cause significant degradation in application performance.
Voltaire’s InfiniBand solutions accelerate application performance. Voltaire offers high-performance, low-
latency solutions that enable applications to reach their full performance potential. Benchmark testing has
found that Voltaire interconnect solutions reduce runtime by as much as 50-300 percent. Voltaire offers high-
performance (10, 20 and 40 Gbps), low-latency (< 2 microseconds) interconnect solutions used in the world’s
highest performance supercomputers and data centers.
InfiniBand is an industry-standard interconnect for high-performance computing (HPC) and enterprise
applications. The combination of high bandwidth, low latency, and scalability with high performance storage
makes InfiniBand the interconnectof choice to power many of the world’s largest and fastest computer systems
and commercial data centers. Voltaire solutions support most major server vendors, operating systems, storage
solutions and chip manufacturers.
1 Gb Ethernet 10 Gb Ethernet Myrinet InfiniBand
Bandwidth 1 Gb/sec 10 Gb/sec 2.5 Gb/sec 10, 20 & 40 Gb/sec
Latency ~10 us 2.5 - 5.5 us < 2 us
Average Efficiency 53% No Entries 68% 74%
Price Per Gig/Port ~$350.00 >~$700.00 ~$225.00 <$100.00
Table 1: Price/performance advantages for InfiniBand
In addition, Voltaire works with leading storage and application vendors to optimize their solutions to alleviate
IPC and file I/O bottlenecks. By combining leading storage technologies with InfiniBand and Voltaire’s Grid
Director™ family of switch products, life sciences organizations can conduct research faster and more
efficiently to gain a clear competitive advantage.
Benchmark
testing has found
that Voltaire
interconnect
solutions reduce
runtime by as
much as 50 – 300%
and increase
application and
file-system
performance 10x .
4
Solution Brief: Higher Education and Research
\
70
60
50
40
30
21
10
00 20 40 60 7010 30 50
# Cores
46.5975% Efficiency
33.0051% Efficiency
Para
llel S
peed
up
Parallel Speedup
Linear Scaling GbE InfiniBand
Figure 1. 75% Parallel Speed Up with Voltaire vs. only 50% with GbE
Building High Performance Clusters
Voltaire offers complete end-to-end server interconnect solutions for speeding life sciences applications. The
three major elements of the solution include:
High-speed, low latency InfiniBand switches �
Fast Storage Access and Scalable File Systems �
Visualization Solutions �
High-performance InfiniBand SwitchesVoltaire’s InfiniBand-based solutions deliver high performance and scalability to compute clusters. Voltaire
offers a complete portfolio of products including a scalable line of InfiniBand switches, high performance I/O
gateways (for seamless connectivity to Ethernet and Fibre Channel networks) and fabric management software.
Voltaire solutions use the Open Fabric Alliance’s OFED drivers and the Open MPI (Message Passing Interface)
libraries to optimize application performance for both MPI-based and socket-based applications.
Figure 2. Voltaire Grid Director 9024 for small-to-medium sized clusters ranging from 16 to 24 nodes
5
For small-to-medium sized clusters, Voltaire offers the Voltaire Grid Director™ 9024. It is a 1U device with
twenty-four 10 Gbps (SDR) or 20 Gbps (DDR) InfiniBand ports. The switch is a high performance, low latency,
fully non-blocking edge or leaf-switch with a throughput of 480 Gbps.
The Grid Director 9024 is well-suited for small InfiniBand fabrics with up to 24 nodes because it includes all of
the necessary management capabilities to function as a stand-alone switch. The Grid Director 9024 is internally
managed and offers comprehensive device and fabric management capabilities. Designed for high-availability
(high MBTF) and easy maintenance, the switch is simple to install and features straightforward initialization.
The solution is scalable as additional switches can be added to support additional nodes.
Figure 3. Voltaire Grid Director 2004 for scalable clusters ranging from 24-96 compute nodes.
For larger clusters ranging from 25–288 compute nodes, Voltaire offers the Grid Director™ 2004 and 2012
multi-service switch—the industry’s highest performing multi-service switches for medium-to-large clusters
and grids. The switch enables high performance non-blocking configurations and features an enterprise-level,
high availability design. The Grid Director 2004 supports up to 96 InfiniBand 4X ports (20 Gbps) and the Grid
Director 2012 supports up to 288 InfiniBand 4X ports (20 Gbps). Voltaire Grid Director switches are scalable
through the use of modular line boards and they feature 10 GbE and Fibre Channel capabilities so the solution
can provide high-performance, integrated SAN and LAN connectivity.
Voltaire has also defined scalable units for deploying larger, scalable clusters. Scalable units are ideal for
constructing large clusters that deliver unparalleled performance to applications. Scalable units combine
compute, interconnect and storage capabilities with scalable file systems.
At the heart of the solution is the Voltaire Grid Director 2012 multi-service switch. Voltaire’s director-class,
multi-service switches offer integrated InfiniBand, GbE and Fibre Channel connectivity in a single chassis. This
enables MPI and storage traffic to run on the same network, a capability that is not available with Ethernet
or proprietary fabrics. By enabling IPC and high performance storage on a single network, Voltaire solutions
enable far greater scalability.
6
Solution Brief: Higher Education and Research
Figure 3. A scalable unit of 200 nodes powered by a Voltaire Grid Director 2012
Scaling out further is made easy by using Voltaire Grid Director switches as core switches to interconnect
multiple scalable units. Such connectivity can be implemented as fully non-blocking or as partially blocking
depending on application requirements or budget constraints.
sFU-8
SMB
SMB
sFB-12sFB-12
sFB-12sFB-12
sFU-8
SMB
SMB
sFB-12sFB-12
sFB-12sFB-12
sFU-8
SMB
SMB
sFB-12sFB-12
sFB-12sFB-12
sFU-8
SMB
SMB
sFB-12sFB-12
sFB-12sFB-12
Figure 4. Multiple scalable units interconnected using a Voltaire Grid Director 2012
7
Fast Storage AccessFor companies looking to incorporate storage into their InfiniBand cluster, Voltaire offers fast I/O capabilities
for storage. Voltaire solutions combine scalable compute and storage capabilities with parallel file systems. By
using InfiniBand with parallel file systems, the server’s CPU overhead is reduced, freeing up CPU cycles for your
application.
ComputeServer
CPUUtilization
Before Now
Software
Network
- Low performance- High overhead on CPU- No scalability
- High performance- CPU available for applications- Scalability to thousands of nodes
IPC Storage
Compute
MPI NFS
ProprietaryInterconnect GbE
IPC Storage
MPI Parallel FileSystem
InfiniBand
At the heart of the solution is the Voltaire Grid Director 2004 multi-service switch (described above). Voltaire’s
director-class, multi- service switches offer seamless InfiniBand, GbE and Fibre Channel connectivity. This
enables MPI and storage traffic to run on the same network, a capability Ethernet and proprietary fabrics do not
offer.
By enabling IPC and high-performance storage on a single network, Voltaire solutions allow companies to leave
behind the limitations of network file systems (NFS) and move to parallel file systems over InfiniBand. This
provides far-greater scalability. Applications can now have effective file I/O rates of 350MB/s compared with
the 50MB/s previously available by using NFS. Additionally, the size of compute clusters is no longer limited by
the limitations imposed by NFS.
Scalable File SystemsRunning scalable file systems over Voltaire InfiniBand solutions creates the most scalable solution in the
industry with more than 1,000 nodes on a single name space, and delivers high performance connectivity for
the storage and client nodes.
Voltaire has significant experience and expertise in enabling large-scale parallel file system deployments.
Such solutions include: Lustre, HP SFS, IBM GPFS, Panasas and PVFS. These solutions, when combined with
InfiniBand, solve two critical problems that NFS creates: limited throughput and limited scalability.
The diagram (Figure 6) below outlines a Voltaire deployment with HP SFS (Lustre) with 1,100 nodes (2,200
cores) all accessing a single file system.
8
Solution Brief: Higher Education and Research
VoltaireISR 9288
Visualization Nodes
Lustre Clients
Compute Nodes Storage Nodes
sCTRL
Voltaire Grid
Multi-PanelDisplays Lustre Servers (OSS, MDS)
DOD T1-05
OSS Sustained FS Performance 900 MB/sClient Side Performance 350 MB/s
Figure 6: Voltaire’s TI-05 installation at the D.O.D – 1,100 nodes on a single file system –
InfiniBand is used for both MPI and Lustre over the same wires.
Visualization SolutionsCustomers in a variety of industry sectors including life sciences, energy, automotive, aerospace, government
and military use Voltaire solutions for visualization on clusters and grids ranging from dozens to hundreds of
nodes.
Visualization clusters typically require very high bandwidth and Voltaire’s InfiniBand solutions with up to
20Gbps per host link are ideal for this. In addition, the low overhead on the CPU frees it to carry out the image
processing faster. The result is a powerful solution that delivers higher levels of resolutions and faster image
processing.
Putting It All Together
Key Features & BenefitsVoltaire solutions for life sciences offer many compelling benefits to users:
High Bandwidth: � Voltaire solutions provide bandwidth of 20 Gbps to allow for faster and more
frequent analysis by life sciences applications.
Lower Latency: � Voltaire solutions provide latency as low as 1.3 microseconds. Moreover, Voltaire’s
InfiniBand-based solutions leverage Remote Direct Memory Access (RDMA) with CPU and OS
bypass-technologies that greatly reduce memory-copy overheads and associated CPU utilization.
Standards-Based: � Voltaire solutions are based on InfiniBand: the only industry-standard, high-
performance interconnect
Flexibility: � Grids and clusters that use Voltaire solutions can be built as a fully non-blocking 20
Gbps fabric or as a lower-bandwidth fabric based on the needs of the application. Moreover, Voltaire
switches are upgradeable in a non-disruptive, hot-pluggable manner.
Figure 7: 3D visualization is
commonly used in
Life Sciences.
9
Fast I/O for Storage: � Voltaire solutions enable parallel file systems over InfiniBand, which offers
far-greater performance and scalability than NFS solutions.
Tested & Certified with Leading ApplicationsBy working closely with leading server and software vendors on integration and testing, Voltaire offers the
fastest and most efficient high-speed interconnect solutions for the life sciences market.
OS Support Linux Enterprise Edition from Novell (SUSE SLES) and Red Hat
(EL/AS)
Supported Parallel File Systems Lustre, HP SFS, IBM GPFS, Panasas, IBRIX, TerraScale’s TeraGrid
Applications Accelrys, Gaussian, AMBER, BLAST, FASTA, GlimmerM, Wise2,
ACT, ClustalW, EMBOSS, HMMER, Image, T-Coffee, Artermis,
CHARMm, Cn3D, GAMESS, GROMACS, RasMol, ReadSeq,
TribeMCL, NAMD, NMRView
Systems & Platform Partners HP, IBM, SUN, NEC, SGI, NEC, Intel, AMD
Customer Success Stories
Customers in a variety of industries that rely on modeling, simulation and analysis leverage Voltaire solutions
for their cluster interconnects. Companies include consumer-product manufacturers, research orgnizations,
pharmaceutical companies, and university development labs throughout the world.
Swiss Institute of BioInformatics (SIB)The Swiss Institute of BioInformatics (SIB) needed a system that could help them speed up discovery of
research for their humanitarian efforts around diseases like mad cow and breast cancer research. They turned
to HP and Voltaire to supply a system that could grow with their needs, but also help speed up the life science
applications running on the cluster.
The Swiss Institute of BioInformatics, through its VitaI IT Computing Institution, created a joint venture
between Oracle, HP and Intel using Voltaire Grid Director™ switches to enable the modeling and analysis.
SIB first built the Vital-IT Computer Center in 2003 using a configuration relying on SAN storage and NFS
services, using gigabit Ethernet for communication among servers. But 18 months after the initial deployment,
SIB realized two major problems.
“With SAN storage, a limited number of servers can be connected to the SAN,” said Dr. Victor Jongeneel of
SIB. “Our servers send and receive data from the compute nodes using NFS, and this turned out to be a major
performance bottleneck. With 64 clients running significant I/O, we needed a faster way than gigabit Ethernet
for all of our compute nodes to share common file space.”
Voltaire solutions support
many leading life sciences
applications.
10
Solution Brief: Higher Education and Research
Storage HP SFS
Test Cluster #1
(8 nodes)
Test Cluster #2
(4 nodes)
Production Cluster
(68 nodes)
Voltaire Grid Director Switch
sFU-8
SMB
SMB
sFB-12sFB-12
sFB-12sFB-12
Figure 8: The system configuration as deployed by Swiss Institute of BioInformatics (SIB)
SIB also has many computing jobs that read large amounts of data into memory, many of which are I/O
bound. Because of the load on the file servers, they would sometimes crash, and jobs could be aborted before
completion. “Because of this, we could not run as many jobs per unit time as we wanted,” Jongeneel said.
The Voltaire interconnect between compute nodes provides much better performance both in terms of
bandwidth and latency,” Jongeneel said. “Computing jobs using HP-MPI, Voltaire MPI and LAM MPI now
run much faster than they used to, and more importantly, our server I/O capacity is no longer a bottleneck for
running any size job.”
SIB selected a configuration of one large cluster consisting of 80 HP Servers with a mix of Itanium2 and
EM64T processors—connected by the Voltaire Grid Director switch. The solution leverages the multi-service
capabilities of Voltaire’s Grid Director switches to enable Fast I/O with storage connectivity. The clusters run
multiple life science applications such as Blast, Gromacs, CHARM and Platform Computing’s scheduler LSF. This
customer is extremely satisfied with the solution because the combination of the Voltaire Grid Director switch,
HP servers, MPI software and File I/O for larger jobs makes for easier file manipulation at high performance.
11
Tokyo Institute of Technology (TiTech)The Tokyo Institute of Technology (TiTech), located in Tokyo, Japan, is one of the leading technical universities
in the world. TiTech has a long history as a world leader in high performance and GRID computing, and houses
one of the world’s largest supercomputers.
Storage
Core
Server
sCTRL sCTRL sCTRL
sCTRL sCTRL
sCTRL sCTRL
sCTRL
The TiTech system serves as an example of how scalable and flexible the Voltaire family of products is. Used
by a wide range of researchers across the university and collaborators across Japan and around the globe, the
solution delivers more than 40 trillion floating point operations per second (TFlops).
The solution tackles computationally difficult problems ranging from:
Analysis of how Avian Flu mutates and is transmitted from birds to humans �
Structural analysis of new materials �
Making buildings more resistant to earthquakes �
More accurate prediction of the earth’s climate �
Visit www.top500.org for more information on this solution.
Contact Voltaire to Learn More
1.800.865.8247 [email protected] www.voltaire.com
©2008 Voltaire Inc. All rights reserved. Voltaire and the Voltaire logo are registered trademarks of Voltaire Inc. Grid Director is a trademark of Voltaire Inc. Other company, product, or service names are the property of their respective owners.
About Voltaire
Voltaire (NASDAQ: VOLT) designs and develops server and storage switching and software solutions that
enable high-performance grid computing within the data center. Voltaire refers to its server and storage
switching and software solutions as the Voltaire Grid Backbone™. Voltaire’s products leverage InfiniBand
technology and include director-class switches, multi-service switches, fixed-port configuration switches,
Ethernet and Fibre Channel routers and standards-based driver and management software. Voltaire’s
solutions have been sold to a wide range of end customers including governmental, research and educational
organizations, as well as market-leading enterprises in the manufacturing, oil and gas, entertainment, life
sciences and financial services industries. More information about Voltaire is available at www.voltaire.com or
by calling 1-800-865-8247.
NoticeReproduction of this publication in any form without prior written permission is not allowed. The information in this publication is subject to change without notice and is provided “AS IS” WITHOUT WARRANTY OF ANY KIND. THE ENTIRE RISK ARISING OUT OF THE USE OR INTERPRETATIONS OF THIS INFORMATION REMAINS WITH RECIPIENT. IN NO EVENT SHALL VOLTAIRE BE LIABLE FOR ANY DIRECT, SPECIAL, PUNITIVE OR OTHER DAMAGES.
Performance results will vary based upon a number of system factors. Some of these include: server configuration of the processor, chip set, memory size, firmware and driver release versions, MPI version and OS kernel version. The configuration or configurations tested or described may or may not be the only available solution. These tests are not a determination of product quality or correctness, nor does it ensure compliance with any federal state or local requirements.
Product names mentioned herein may be trademarks and/or registered trademarks of their respective companies.