High Performan Ce Computing: Making Everythi NG Easier!

g Easier!
Making Everythin
ial Edition
Sun and AMD Spec
e
c
n
a
m
r
o
f
r
e
High P
Computing
Learn to:
Pick out hardware and software
Find the best vendor to work with
Get your people up to speed on
HPC
Douglas Eadline, PhD
High Performance
Computing
FOR
DUMmIES
SUN AND AMD SPECIAL EDITION
by Douglas Eadline, PhD
High Performance Computing For Dummies, Sun and AMD Special Edition
Published by
Wiley Publishing, Inc.
111 River Street
Hoboken, NJ 07030-5774
Copyright 2009 by Wiley Publishing, Inc., Indianapolis, Indiana
Published by Wiley Publishing, Inc., Indianapolis, Indiana
No part of this publication may be reproduced, stored in a retrieval system or transmitted in any
form or by any means, electronic, mechanical, photocopying, recording, scanning or otherwise,
except as permitted under Sections 107 or 108 of the 1976 United States Copyright Act, without the
prior written permission of the Publisher. Requests to the Publisher for permission should be
addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ
07030, (201) 748-6011, fax (201) 748-6008, or online at http://www.wiley.com/go/permissions.
Trademarks: Wiley, the Wiley Publishing logo, For Dummies, the Dummies Man logo, A Reference
for the Rest of Us!, The Dummies Way, Dummies.com, Making Everything Easier, and related trade
dress are trademarks or registered trademarks of John Wiley & Sons, Inc. and/or its affiliates in the
United States and other countries, and may not be used without written permission. AMD, the AMD
Arrow logo, AMD Opteron, AMD Virtualization, AMD-V, and combinations thereof are registered
trademarks of Advanced Micro Devices, Inc. HyperTransport is a licensed trademark of the
HyperTransport Technology Consortium. Sun, the Sun logo, Solaris, StorageTek, Sun Fire, Sun xVM
Ops Center, and Sun Ray are trademarks or registered trademarks of Sun Microsystems, Inc. in the
United States and other countries. All other trademarks are the property of their respective owners.
Wiley Publishing, Inc., is not associated with any product or vendor mentioned in this book.
LIMIT OF LIABILITY/DISCLAIMER OF WARRANTY: THE PUBLISHER AND THE AUTHOR MAKE
NO REPRESENTATIONS OR WARRANTIES WITH RESPECT TO THE ACCURACY OR COMPLETENESS OF THE CONTENTS OF THIS WORK AND SPECIFICALLY DISCLAIM ALL WARRANTIES,
INCLUDING WITHOUT LIMITATION WARRANTIES OF FITNESS FOR A PARTICULAR PURPOSE.
NO WARRANTY MAY BE CREATED OR EXTENDED BY SALES OR PROMOTIONAL MATERIALS.
THE ADVICE AND STRATEGIES CONTAINED HEREIN MAY NOT BE SUITABLE FOR EVERY SITUATION. THIS WORK IS SOLD WITH THE UNDERSTANDING THAT THE PUBLISHER IS NOT
ENGAGED IN RENDERING LEGAL, ACCOUNTING, OR OTHER PROFESSIONAL SERVICES. IF PROFESSIONAL ASSISTANCE IS REQUIRED, THE SERVICES OF A COMPETENT PROFESSIONAL
PERSON SHOULD BE SOUGHT. NEITHER THE PUBLISHER NOR THE AUTHOR SHALL BE LIABLE
FOR DAMAGES ARISING HEREFROM. THE FACT THAT AN ORGANIZATION OR WEBSITE IS
REFERRED TO IN THIS WORK AS A CITATION AND/OR A POTENTIAL SOURCE OF FURTHER
INFORMATION DOES NOT MEAN THAT THE AUTHOR OR THE PUBLISHER ENDORSES THE
INFORMATION THE ORGANIZATION OR WEBSITE MAY PROVIDE OR RECOMMENDATIONS IT
MAY MAKE. FURTHER, READERS SHOULD BE AWARE THAT INTERNET WEBSITES LISTED IN
THIS WORK MAY HAVE CHANGED OR DISAPPEARED BETWEEN WHEN THIS WORK WAS WRITTEN AND WHEN IT IS READ.
For general information on our other products and services, please contact our Customer Care
Department within the U.S. at 877-762-2974, outside the U.S. at 317-572-3993, or fax 317-572-4002. For
details on how to create a custom For Dummies book for your business or organization, contact
bizdev@wiley.com. For information about licensing the For Dummies brand for products or services, contact BrandedRights&Licenses@Wiley.com.
ISBN: 978-0-470-49008-2
Manufactured in the United States of America
10 9 8 7 6 5 4 3 2 1
Publishers Acknowledgments
Project Editor: Jennifer Bingham
Editorial Manager: Rev Mengle
Sun Contributors: Tony Warner, Patrice Brancato, Allison Michlig, Frances Sun
AMD Contributors: Jeff Jones, Jeff Underhill
Table of Contents
Introduction....................................................... 1
About This Book ........................................................................ 1
Icons Used in This Book ............................................................ 2
Chapter 1: HPC: Its Not Just for Rocket

Scientists Any More ........................................... 3
The Forms of HPC ..................................................................... 4
Who Uses HPC Today? .............................................................. 6
Who Should Be Using HPC? ...................................................... 8
Chapter 2: Getting to HPC ................................. 9

Enter the Commodity Cluster................................................... 9
Its About Choice ...................................................................... 11
What Does a Cluster Look Like? ............................................. 12
If You Need Speed ................................................................... 16
Of Cores, Threads, and Nodes ............................................... 19
What about the Data? .............................................................. 20
Chapter 3: HPC Building Blocks ....................... 21

Choosing Cluster Hardware.................................................... 21
Crunching numbers: Processors and nodes .............. 21
A little help: Co-processors .......................................... 23
The communication: Interconnects ........................... 23
Remembering the storage ........................................... 25
Racking and stacking ..................................................... 27
Power and cooling ......................................................... 27
Finding Cluster Software ......................................................... 27
Operating systems ........................................................ 28
HPC glue: Cluster software ........................................... 28
File systems .................................................................... 30
Sharing is caring: HPC resource schedulers .............. 31
Ready to run application software ............................. 32
Provisioning: Creating the cluster ............................... 32
Cluster Tool Kits ...................................................................... 33
iv
High Performance Computing For Dummies
Chapter 4: Pulling It All Together ..................... 35

Following the HPC Recipe ....................................................... 35
Start with the personnel ............................................... 35
Consider the software ................................................... 36
Specify the hardware .................................................... 36
Who Is Going To Stand By Me? .................................... 37
Following Some Dos and Donts for HPC Cluster
Procurement ......................................................................... 39
Understanding Benchmarks ................................................... 41
Chapter 5: The HPC Community ....................... 43

Understanding Open Source .................................................. 43
Looking at How HPC Works with Open Source .................... 44
Chapter 6: Ten (Okay, Six) HPC Takeaways ...... 45
Introduction
he quest to capture and manipulate the world in digital form

began when the first vacuum tube computers were used
to calculate trajectories and crack codes. Viewed through
todays technology, these problems seem trivial and could
easily be solved on your cell phone and PDA. The quest to
capture and compute continues to this day. The power to
model and manipulate our world in silicon has enabled vast
changes in how we conduct science, business, and even our
everyday lives. From the next scientific breakthrough to new
and better products to a greener world, High Performance
Computing (HPC) is playing a central role in all these efforts.
In simple terms, HPC enables us to first model then manipulate those things that are important to us. HPC changes everything. It is too important to ignore or push aside. Indeed, HPC
has moved from a selective and expensive endeavor to a
cost-effective technology within reach of virtually every
budget. This book will help you to get a handle on exactly
what HPC does and can be.
High Performance Computing For Dummies, Sun and AMD
Special Edition is intended for anyone who has heard about
the many benefits of using HPC (such as streamlining processes or saving money). This book explains what HPC is and
shows how it can help you or others within your company.
About This Book

HPC is a powerful technique, but it is not available at the corner
office supply center (yet) nor can it solve every product or
process problem you have. It requires a dedicated effort and
commitment to new and exciting technologies. Finding the
right partners and technologies is critically important. Much
of this book is designed to help with this process. Be forewarned:
Results may not be instantaneous and it could take more than

a few business quarters before your efforts pay off. However,
there are plenty of HPC success stories out there, and perhaps
yours is next.
It may be hard to imagine, but were at the beginning of the
High Performance Computing era. This book is an attempt to
give you a high level snapshot where things are today. Trying
to predict what will happen next year or even next week is
anyones guess. Right now, there may be some research group
designing a new product or on the verge of a great discovery
because they can hold and manipulate objects in the digital
domain. HPC will undoubtedly lead to future breakthroughs.
Along with these successes, expect that there will be novel
ways in which to derive a business advantage using HPC.
Dont be afraid to let your mind wander.
This book was sponsored by and written in collaboration with
Sun and AMD.
Icons Used in This Book

Throughout the margins of this book, you find several helpful
little icons that guide you to certain types of information:
Yes, I know. The whole book is full of technical stuff. But this
technical stuff tends to be a little off the beaten track. You
dont have to read it to follow along, but if you do, you earn
extra credit.
This icon alerts you to key concepts that you might want to
commit to memory.
Flagging a tip is just a nice way to warn you that if youre not
careful, you may end up overlooking something important
that can save you time or money.
Chapter 1
HPC: Its Not Just

for Rocket Scientists
Any More
In This Chapter
Looking at the forms of HPC
Examining who uses HPC today
Explaining who ought to be using it
ention the word supercomputer to someone and they

automatically think of monstrously complicated
machines solving problems no one really understands. Maybe
they think of flashing lights and some super intelligence that can
beat humans at chess or figure out the meaning of life, the universe, and everything. (Its 42 for those who are interested.)
Back in the day, this was not an altogether untrue view of
supercomputing. With an entry fee of at least seven figures,
supercomputing was for the serious scientists and engineers
who needed to crunch numbers as fast as possible.
Today we have a different world. The custom supercomputer of
yesteryear has given way to commodity-based supercomputing,
or what is now called High Performance Computing (HPC). In
todays HPC world, it is not uncommon for the supercomputer to
use the same hardware found in Web servers and even desktop
workstations. The HPC world is now open to almost everyone
because the cost of entry is at an all-time low.
HPC mud flaps

HPC can show up in the most unexpected places like under a truck.
Large trucks spend a lot of time
moving. Pushing against the air
while driving takes energy (in this
case gasoline) and energy costs
money. Almost all trucks have some
kind of mud flaps to prevent road dirt
and debris from hitting the truck. A
midsized truck maker wondered how
much it cost to push those mud flaps
through the air. Using HPC, they were
able to determine that trimming and

tapering the flaps can cut about $400
from a typical trucks annual gas bill.
Although it might not sound like much
for a single truck, when you have a
fleet of 1,000 trucks, those savings
add up very quickly. The company
has now moved far beyond mud flaps
and is using HPC to help increase the
efficiency of their truck designs, thus
saving even more money.
To many organizations, HPC is now considered an essential

part of business success. Your competition may be using HPC
right now. They wont talk much about it because its considered
a competitive advantage. Of one thing you can be sure, however;
theyre designing new products, optimizing manufacturing and
delivery processes, solving production problems, mining data,
and simulating everything from business process to shipping
crates all in an effort to become more competitive, profitable,
and green. HPC may very well be the new secret weapon.
The Forms of HPC

Ask ten people what HPC is and you will get ten different answers.
Applying compute cycles to a problem can be done in many
ways. Currently, there seem to be four modes in which you
can obtain the cycles needed for typical HPC problems:
The commodity HPC cluster: Over the last ten years, the
HPC cluster has disrupted the entire supercomputing
market. Built from standard off-the-shelf servers and high
speed interconnects, a typical HPC system can deliver
industry-leading, cost-effective performance. A typical
cluster can employ hundreds, thousands, and even tens
Chapter 1: HPC: Its Not Just for Rocket Scientists Any More
of thousands of servers all working together on a single

problem (this is the high tech equivalent of a divide and
conquer approach to solving large problems). Because
of high performance and low cost, the commodity cluster
is by far the most popular form of HPC computing. Also
keep in mind the compatibility advantage x86 commodity servers are ubiquitous.
Dedicated supercomputer: In the past, the dedicated
supercomputer was the only way to throw a large number
of compute cycles at a problem. Supercomputers are still
produced today and often use specialized non-commodity
components. Depending on your needs, the supercomputer
may be the best solution although it doesnt offer the
commodity price advantage.
The Top500
No, the Top500 isnt a car race. It is
a list of the worlds fastest computers. Of course, some background is
needed. To get on the Top500 list,
you must run a single benchmark
program on your HPC system and
submit it to the Top500 organization.
The list is created twice a year and
includes some rather large systems.
Not all Top500 systems are clusters,
but many of them are built from the
same technology. Of course, as with
all lists, there are some limitations.
First, the list is for a single benchmark
(HPL or High Performance Linpack).
Results for other benchmarks may
shuffle the standings, but systems
are likely to remain in roughly the
same place in the list if other benchmarks are used. Second, the list only
includes those systems that were
submitted for consideration. There
may be HPC systems out there that

are proprietary or not interested in
the Top500 ranking.
Yet despite these limitations, the
Top500 list is the wealth of historical
data. The list was started in 1993 and
has data on vendors, organizations,
processors, memory, and so on for
each entry in the list. You can view
the list (past and present) by going to
http://top500.org/.
AMD and Sun Microsystems have
many accomplishments on the
Top500 list. On the most recent version of the list, seven of the top ten
systems on the list use AMD technology. The Texas Advanced Computing
Center is home to the sixth fastest system in the world (based on
the current list) called Ranger. See
Chapter 2 for more on this system.

HPC cloud computing: This method is relatively new and
employs the Internet as a basis for a cycles-as-a-service
model of computing. The compute cycles in question live
in the cloud somewhere allowing a user to request remote
access to cycles on-demand. An HPC cloud provides
dynamic and scalable resources (and possibly virtualization) to the end-user as a service. Although clouds can be
cost effective and allow HPC to be purchased as an expense
and not a capital asset, it also places some layers between
the user and hardware that may reduce performance.
Grid computing: Grid is similar to cloud computing,
but requires more control by the end-user. Its main use
is academic projects where local HPC clusters are connected and shared on a national and international level.
Some computational grids span the globe while others
are located within a single organization.
Who Uses HPC Today?

The worldwide HPC market is growing rapidly. According to IDC,
the total HPC market was $10 billion in 2007 and is expected
to hit $11.7 billion around 2012. In terms of market share, IDC
defines price bands by system size as shown in Table 1-1.
Table 1-1
2007 HPC Market Share as

Function of System Cost
System Size
Market Share
Supercomputers (Over $500K)
$2.7B
Technical Divisional
($250K$500K)
$1.6B
Technical Departmental
($100K$250K)
$3.4B
Technical Workgroup
(under $100K)
$2.4B
Successful HPC applications span many industrial, government,

and academic sectors. The following is a list of major areas
where HPC has a significant presence:
Bio-sciences and the human genome: Drug discovery,
disease detection/prevention
Chapter 1: HPC: Its Not Just for Rocket Scientists Any More
Computer aided engineering (CAE): Automotive design

and testing, transportation, structural, mechanical design
Chemical engineering: Process and molecular design
Digital content creation (DCC) and distribution:
Computer aided graphics in film and media
Economics/financial: Wall Street risk analysis, portfolio
management, automated trading
Electronic design and automation (EDA): Electronic
component design and verification
Geosciences and geo-engineering: Oil and gas exploration
and reservoir modeling
Mechanical design and drafting: 2D and 3D design and
verification, mechanical modeling
Defense and energy: Nuclear stewardship, basic and
applied research
Government labs: Basic and applied research
University/academic: Basic and applied research
Weather forecasting: Near term and climate/earth
modeling
Cultivating the HPC edge

Interested in learning more about
using HPC as a competitive tool? You
arent alone. As a mater of fact, the
Council on Competitiveness (CoC)
wants to help you! Who is the CoC?
Glad you asked, theyre a group of
corporate CEOs, university presidents, and labor leaders committed to
the future prosperity of all Americans
and enhanced U.S. competitiveness
in the global economy through the
creation of high-value economic
activity in the United States. In other
words, they want to help your business succeed.
The CoC is a nonpartisan, nongovernmental organization based in

Washington, D.C. They shape the
debate on competitiveness by bringing together business, labor, academic, and government leaders to
evaluate economic challenges and
opportunities. The High Performance
Computing (HPC) Initiative is intended
to stimulate and facilitate wider usage
of HPC across the private sector to
propel productivity, innovation, and
competitiveness. For more information, case studies, and surveys check
out www.compete.org.

The list could be longer as more and more areas are finding HPC
useful as a tool to better understand their science, market,
products, and customers. As pioneers, the government and
academic sectors have been successfully using and validating
HPC methods for well over a decade. The secret is out and the
same advantages enjoyed by leading researchers can be had
by almost anyone.
Who Should Be Using HPC?

The easy answer to this question is: anyone interested in
faster times to solution, better science, informed decisions,
more competitive products, and how those things can drive
profits higher. High Performance Computing represents a tremendous competitive edge in the marketplace because it can
give users the ability to quickly model and then manipulate
a product or process to see the impact of various decisions
before they are made. Consider a car as it moves through the
air. Once it has been modeled in an HPC system, the end-user
can look at the air flow at points that would be almost impossible with a physical model in a wind tunnel.
The possibilities are endless. Existing products and processes
can be optimized to reduce cost, and new avenues of development can be explored at lower cost and with faster times to
market. The question to ask is What information would allow
you to make better decisions and products? Chances are
HPC can help with the answer.
Of coffee, chips, and soap

HPC has made huge inroads in the
government and educational sectors.
Perhaps one of the leading industrial champions of HPC is Procter &
Gamble (P&G). This fact sometimes
surprises people because P&G is
often seen as a low-tech supplier
of household products. Quite the
opposite is true. P&G is one of the
pioneers of industrial HPC. They
can boast of significant cost savings

and innovation across many product lines, including Folgers coffee,
Pringles potato chips, Tide detergent,
and many more. Unlike its competitors, P&G has been very vocal about
its success with HPC. And, according to P&G, the successes have been
well worth the investment.
Chapter 2
Getting to HPC
In This Chapter
Examining the commodity cluster
Looking at whats available
Getting to know the cluster
Drilling down into some details
n the past when you wanted to deploy a supercomputer,

you picked up the phone and said, Hello Supercomputer
Company, I want one of your model Z3 systems, in blue. Sure
enough, in a few weeks (or months) you were the proud owner
of a new supercomputer (complete with a seven figure price tag).
There was a good chance the company you bought it from would
teach your personnel how to use it. There would be manuals to
assist users and administrators as well. When something needed
to be changed or updated it was tested before it was turned over
to your system administrators. There was even a support line
where you could get questions answered about the system.
Once upon a time, this was the face of supercomputing. The
entire system was integrated from top to bottom by a single
company. From a hardware standpoint, supercomputers
employed a handful of specialized processors that were
designed to quickly operate on batches of floating point numbers.
So how did HPC move from something that was once confined
to the elite of computing to that which almost any organization
can now cost effectively deploy?
Enter the Commodity Cluster

As nice as the integrated supercomputer was, it did have
some drawbacks. For one, the price was rather steep seven
10

figures. There was no doubt that leading edge required high
cost. The high cost also reflected the significant amount of
custom engineering required for these systems. The extreme
performance came from specialized processors and memory
systems configured in such a way as to move numerical calculations through a vector pipeline as fast as possible.
As supercomputer system performance increased, so did the
cost of designing and fabricating custom processors and other
components. At the same time, the commodity computer
market was experiencing rapid growth and maturation of x86
processors. Often considered office processors, no one really
considered x86s to be good at floating point calculations. This
presumption began to change as floating point units were integrated into x86 processors and each new generation delivered
significant incremental performance
Today its common for each commodity system based on x86
processors to actually contain multiple processors combined
into a single physical processor. When multiple processors
are integrated into a single physical processor, the individual
processors are referred to as cores, hence the buzz about
multi-core processors these days.
In addition to getting faster, supercomputer-specific processors
also got more expensive to manufacture. Indeed, the expense
became so high that it was no longer feasible to design and
fabricate supercomputer-specific processors. In order to earn
a profit, processors have to be sold by the tens of millions.
Supercomputers and even some UNIX workstations could
not justify this expense. In addition, for certain problems it
was found that by breaking a problem up into smaller subproblems, collections of commodity processors could achieve
performance levels that could rival those of their supercomputing kin. Indeed, as commodity processors from companies
like AMD improved, collections (or clusters) of servers (each
server in a cluster is commonly referred to as a node) began
to outpace even the fastest supercomputers. The move to
commodity hardware had begun.
At the same time, there was a small but growing open source
software movement afoot. Many of the software tools needed
to support a UNIX-like environment were freely available under
the GNU software license. The final piece of the software puzzle
was the advent of the Linux operating system. Linux was
Chapter 2: Getting to HPC
11
essentially a plug-and-play replacement for UNIX the then

de facto supercomputing operating system. In addition, Linux
didnt require a license fee and wasnt encumbered by restrictive
intellectual property agreements.
Another enabler of the commodity cluster was the Internet.
The Internet helped in two ways. First, it ignited the demand
for Web servers whose compact data center design could offer
a low-cost building block for commodity HPC clusters. And
second, it opened up an avenue for international collaboration,
helping solve many of the early HPC issues. Entire communities
developed around various approaches to commodity-based
HPC. Internet access to open software projects fueled this rapid
growth and adoption.
By the turn of the millennium, the cost of entry to the supercomputing market had gone from seven figures to the low five
figures. More people could now enter the game and have access
to a level of computing previously reserved for the technical
elite. Typical price-to-performance ratios, key metrics of the
worth of an HPC system, changed as well. Many early adopters
were reporting at least a factor-of-ten improvement of price-toperformance. Thus, many more problems could be explored,
understood, and ultimately solved using HPC methods.
Its About Choice

As HPC cluster solutions began to appear, one thing became
painfully obvious. The components required to create a commodity cluster would likely come from many different sources,
without a computing company taking responsibility for the
whole cluster. For example, the compute nodes could come
from one vendor, and the interconnect from another. If you
added a large storage component, then yet another vendor
could be attending your cluster party. The software was just
as scattered. There was a multitude of open software packages
available to HPC users. Popular Linux distributions included
many of the necessary software packages; however, successful
HPC systems often required some blend of specialized open
source and commercial packages.
Such variety can be both good and bad. On the good side,
users can control what is (and is not) bought, and can optimize HPC investments for their specific needs. Unlike the
12

large monolithic supercomputers, theres little need to buy
technology you dont use or want. In addition, the price pressures from the commodity market ensure that hardware is
both readily available and cost effective. In other words, compared to the days of old, commodity HPC is something of a
bargain!
The other side of the coin is not so popular, however. Elite
users have no problem specifying and designing HPC clusters,
but those who are not as technically adroit are not as lucky.
The right hardware and software choices could mean the difference between success and failure.
Getting it right is often seen as an impediment to cluster adoption. In addition to design, there is also the interoperability
issue. There are a handful of vendors who specialize in cluster
design and construction; however, integrating a large amount
of hardware and software can lead to compatibility issues.
Because the integrators are often not the source of the hardware and software components, identifying the appropriate
person for problem ownership and resolution can be tricky.
For this reason, choosing the right integrator/vendor with a
good vertical product line can be very important.
Choosing the right vendor (with the right products) ensures
that your HPC system just works. This is more important
today then ever before. Processors and networking are getting faster all the time. Its important to ensure that all the
components of your environment work together and no one
area becomes a bottleneck
What Does a Cluster Look Like?

Trying to pin down what a cluster actually looks like is rather
difficult. The nature of HPC cluster computing mandates that
users consider a solution for their specific application needs.
Since your needs are likely to be different from everyone
elses, it is very likely that your cluster will be configured differently than everyone elses. The differences can be small or
large depending on your application and users needs.
There are, however, some common features that all clusters
share. The first is a head or master node. This node is normally a gateway to a shared network where users log in
13
to the cluster. The head node has one or more networks with
which it communicates to the worker (or compute) nodes.
These networks are private and are only accessible inside
the cluster.
All clusters have worker nodes that do the bulk of the computing. These nodes are almost always identical throughout
the cluster (although there is no strict requirement that all
nodes need to be the same). Indeed in some cases it is advisable to configure your cluster with different compute nodes.
This all depends on what applications you are trying to run.
For example some applications require more local memory
than others, while some require a specific processor architecture in order to run at best performance.
In terms of networks, a cluster may have as little as one private
network, which is usually Gigabit Ethernet (GigE), although
InfiniBand is becoming a popular option. Almost all servers
have at least two GigE connections on the motherboard and
therefore all clusters have at least one private GigE network. A
cluster will often generate three types of traffic:
Computation traffic between compute nodes.
File system traffic often from an NFS (Network File
System) server. (But not always. Direct attach storage
has its advantages.)
Administration traffic that provides node monitoring and
job control across the cluster.
Depending on your applications, compute and/or file system
traffic may dominate the cluster network and cause the nodes
to become idle. For this reason, additional networks are
added to increase overall cluster network throughput. In
many cases, a high speed interconnect is used for the second
network. Currently the two most popular choices for this
interconnect are 10 Gigabit Ethernet (10GigE) and InfiniBand.
As a lower-cost solution it is possible to use a second GigE
network as well, however this would not offer anywhere near
the network performance of 10GigE or InfiniBand.
Figure 2-1 shows a typical cluster configuration. Theres a
head node that may contain a large amount of storage that is
shared via the network by all the worker nodes (using some
form of Network File System). A varying number of worker
14

nodes communicate with each other and the head node over
the available private networks. The number of worker nodes
can be quite large. For example the Sun Constellation Linux
Cluster called Ranger, based on AMD Opteron processors
and located at the Texas Advanced Computing Center, has
3,936 nodes, each containing four Quad-core AMD Opteron
processors for a total of 62,976 compute cores (see Figure 2-2).
Blade nodes are connected using InfiniBand and a Sun Data
Center Switch 3456 (a high performance 3456 port InfiniBand
switch).
10 GigE, InfiniBand
Compute Traffic
Head
disk
NFS
to local
Worker Worker
Node Node
Worker Worker
Node Node
Gigabit Ethernet Switch

File
System/Administration
Traffic
Figure 2-1: A typical cluster configuration.
There are a few variations to be considered with any cluster

design. The head node may be broken into several (or many)
nodes. There can be multiple file serving nodes, multiple user
login nodes, and separate administration nodes.
A real live cluster often looks like that shown in Figure 2-2. The
nodes are placed in standard equipment racks and usually
require a data center environment (raised floor, chilled air, and
so on).
Compute nodes or servers are placed in the equipment racks. A
typical 1U server is shown in Figure 2-3. Note that the type and
number of processors, amount of memory, and even the presence of a local storage device are all determined by the user.
15
Image courtesy of the Texas Advanced Computing Center and Advanced Micro Devices.
Figure 2-2: Ranger provides computational capabilities to the national

research community and is built using the Sun Constellation
System architecture from Sun Microsystems.
Figure 2-3: Typical 1U Cluster Node the Sun Fire X2200 M2 server
(1U indicates that the server is 1.75 inches high)
Currently, there is a trend to move away from individual 1U

pizza box nodes to the increasingly popular bladed systems
like those shown in Figure 2-4. A blade system allows for
shared/redundant power and cooling and management while
improving the node density. A blade-based cluster can also
reduce maintenance costs and improve reliability.
Clusters are often broken into two categories: capability and
capacity. A capability cluster is designed to handle (or be
capable of handling) large compute jobs that may employ
every node in the cluster. Capacity clusters, on the other
hand, are those that are used to deliver a certain amount of
16

computing capacity to the end users. For instance, a capacity
cluster may support hundreds of users running any number
of jobs that require a smaller number of nodes. Most clusters
are used in a capacity fashion.
Figure 2-4: The Sun Blade 6048 chassis holds up to 48 blade server modules, 1,152 cores, delivering up to 12 TFLOPS in a single rack.
If You Need Speed

High performance clusters are used where time to solution is
important. They are also used in cases where a problem is so
big it cant fit on one single computer. To increase computing throughput, HPC clusters are used in a variety of ways.
The first and easiest way is to allow the cluster to act as a
compute farm. Instead of running a job on a local workstation,
it is submitted to the cluster for execution. The cluster will
manage the resources needed for the job and assign the job
to a work queue. When the resource (for instance, a server) is
17
available, the job gets executed and the results are returned
to the user. Users who need to run many similar jobs with
different parameters or data sets find clusters ideal for this
kind of work. They can submit hundreds of jobs and allow the
cluster to manage the work flow. Depending on the resources,
all the jobs may run at the same time or some may wait in the
queue while other jobs finish. This type of computing is local
to a cluster node, which means the node doesnt communicate
with other nodes, but may need high speed file system access.
An example of this usage mode is given in Figure 2-5. Each
job can be from a different user or the same user. The jobs
can run concurrently, thus increasing computing throughput.
Even though more jobs can be run in the same time period,
they never run any faster than normal. Because each job is
independent, there is no communication between the jobs.
Many
Single
Core
Programs
Cluster
Results
Figure 2-5: A cluster used to simultaneously run many core independent
jobs (each small square box is a cluster node).
Another way to use clusters is to break large HPC jobs up into

smaller sub-jobs and run each sub-job on different nodes.
This process is done at the software level and, thus, programs
18

may need to be changed to run in this fashion. The most
popular method for creating cluster programs is to use the
MPI (Message Passing Interface) library and insert communication calls in the program. Because these sub-jobs need to
communicate with each other, they can create a large amount
of compute traffic within the cluster. For this reason, high performance networks are used to handle this level of traffic. It
is not uncommon to use many hundreds of nodes for a single
program. Again, the program must request the resources (for
example, I need 100 processors) from the cluster and when
the resources are ready, the cluster will run the program. This
type of program is called parallel because it uses many cores
to collaborate toward solving the same problem. An example
of a parallel program running on a cluster is given in Figure
2-6. The program employs all the processors and must communicate with the other nodes. The parallel program will run
much faster than using a single node; however, the programmer must explicitly create a parallel program.
One Parallel
Program
Cluster
Result
Figure 2-6: A cluster used to run a parallel program. (Each small square box
is a cluster node).
19
Both modes of computation, compute farm and parallel, can

operate on a cluster at the same time. Of course if a parallel
program needs all the nodes, nothing else can run, but this is
not often the case. There is usually scheduling software that
handles the allocation policies for the cluster.
Breaking up a single program into parts to run across multiple
cores is a rather obvious way to increase speed. Like building a house, the more workers you have, the faster you can
complete the job. Some parts of the construction are highly
parallel, like framing the wall, painting, applying siding, and
so on. Other steps may depend on one worker and cause the
others to wait. For instance, you cant finish the walls until the
electrical work is done. Thus, there are limits to the number
of workers you can add before there is no increase in speed at
which you build a house.
Parallel computing works in exactly the same way. Computer
programs normally have slow steps that limit how fast they
can actually run. This effect is called Amdahls law (see the
sidebar of the same name) and it essentially tells you how
much speed-up you can expect from a given computer program when it is run across multiple cores.
Of Cores, Threads, and Nodes

When discussing HPC, terms like cores, threads, and nodes often
get used in very informal ways. In this book, I refer to cores
as processing units within a modern day processor. A typical
motherboard may hold 1, 2, or 4 processors and is contained
in a node. Thus, the total cores per node is a function of how
many processors are present and how many cores are in each
multi-core processor. A typical node may have 2 processors,
each with 6 cores for a total of 12 cores per node. (A node can
also be defined as one operating system instance.)
From a software standpoint, programs on a multi-core node
are run in a Symmetric Multi-Processing (SMP) mode. This
designation means that multiple programs (or processes) can
be running at the same time on various cores. It also allows
for a single program to use multiple cores by pulling the program apart into threads. Threaded programming is parallel
programming designed for a single SMP node. (It is possible,
but not advisable, to run threads between nodes unless
20

you have specialized hardware and software.) A threaded
program thus starts out as a single program but soon breaks
itself into separate but highly connected parts to take advantage of the multiple cores in the SMP motherboard node.
Thread programming for HPC can be accomplished using the
OpenMP model supported by almost all compilers. Like
threads, OpenMP isnt designed for off node operation,
whereas MPI is. A basic rule is to use MPI between nodes and
OpenMP/threads within a single node. And, keep in mind parallel MPI programs will run on SMP nodes as well. Finally, it is possible to create hybrid programs that use both OpenMP and MPI!
What about the Data?

One final note about parallel computing. In addition to the parallel computation shown in Figure 2-6, there may also be a need
for parallel data access. Imagine if all the nodes in Figure 2-6
were required to read and write data from the same file at the
same time. A bottleneck would develop and limit the speed of
your program. For this reason, parallel file systems have been
developed. These solutions often involve additional hardware
and software. I take a closer look at this issue in Chapter 4.
Amdahls law
In the early days of parallel computing, Gene Amdahl threw a bucket of
cold water on those who got overly
excited about piling on processors.
Amdahl showed you cant keep
throwing more and more cores at a
problem and expect it to keep speeding up. His law works as follows. The
parallel speedup of any program is
limited by the time needed for any
sequential portions of the program to
be completed. For example, if a program runs for ten hours on a single

core and has a sequential (nonparallelizable) portion that takes
one hour, then no matter how many
cores you devote to the program it
will never go faster than one hour.
Seems obvious and it is. Amdahls
law affects the scalability of a
program as well.
Chapter 3
HPC Building Blocks

In This Chapter
Getting the best hardware for the job
Making sure your software fits the bill
Looking at cluster tool kits
PC systems are built from many components. There

are some common elements, but clusters may employ a
wide range of hardware and software solutions. This chapter
goes over some of the important cluster components.
Choosing Cluster Hardware

Obviously if youre looking into HPC, youre going to have to
make sure you get hardware that can handle it! This section
goes over what you need to make sure you have in place in
order to set up an HPC system.
Crunching numbers:
Processors and nodes
The processor is the workhorse of the cluster. And, keeping the
workhorse busy is the key to good performance. Parallel programs are often distributed across many nodes of the cluster
(for more info on this, see Chapter 2). However, multi-core has
changed this situation a bit. Cluster nodes may now have 8 or
even 16 cores per node (for example, the Sun Fire X4440 server
with four Quad-Core AMD Opteron processors). It is
22

possible for whole HPC applications to fit on a single cluster
node. (Sometimes this helps performance and sometimes it
hurts performance it all depends).
The choice of processor is very important because cluster
installations often rely on scalable processor performance.
Advances in the x86 architecture such as simultaneous 32/64
bit operation (as can be found in AMD64 technology), integrated
memory controllers (as can be found in AMD Opteron processors), and technologies similar to AMD HyperTransport
technology have propelled commodity processors to the forefront of HPC. As an example, in a recent cluster procurement,
Lawrence Livermore, Los Alamos, and Sandia National Labs all
chose AMD Opteron processor-based clusters. As far as power
users go, those government guys like to think big. They deployed
a total of 48,384 cores as part of the project. That would be
12,096 Quad-Core AMD Opteron processors for a combined
processing power of up to 438 TFLOPS. Recently, the Jaguar
system at Oak Ridge National Lab has exceeded 1 PFLOP using
Quad-Core AMD Opteron processors. (See the What the flip are
FLOPS? sidebar.)
Depending on the design of the cluster, nodes can be fat (lots
of cores, disk, and memory), thin (small number of cores and
memory), or anything in between. Some applications work well
in either type of node while others work best with a particular
configuration.
Pay attention to the amount of memory. In general, more cores
per node means more memory per node because each core
could be asked to run a totally separate program. Many
HPC cores require a large amount of memory. Check your
application(s) memory requirements and size your nodes
appropriately. This is something sometimes overlooked when
upgrading from dual- to quad-core processors if you dont
increase the memory capacity, you effectively reduce the
amount of memory per core, which could in turn limit some of
the advantages from upgrading in the first place.
Another feature to consider for nodes is system management.
The Intelligent Platform Management Interface (IPMI) specification defines a set of common interfaces to a computer system.
System administrators can use IPMI to monitor system health
and manage the system. Often IPMI features can be managed
over a network interface, which is very important with large
numbers of nodes often located in a remote data center.
Chapter 3: HPC Building Blocks
23
A little help: Co-processors

Back in the supercomputer days, there were products called
array processors that were designed to do certain mathematic
operations very quickly. Recently, there has been a resurgence of this concept through the use of GP-GPU (General
Purpose Graphical Processing Units) or as they are sometimes called, video cards. Interestingly, this trend is not new in
HPC, where it is not uncommon for practitioners to use new
technologies in innovative ways to solve complex problems.
As with array processors, these devices can accelerate certain types of mathematical operations. In the case of GP-GPUs,
however, dual-use as a commodity product has made this a
relatively low-cost solution. One word of advice, take note
of the precision and FLOPS reported for real applications
and remember existing parallel applications must be reprogrammed to use these types of devices. There are efforts
to help address this problem. For instance, the recently
approved OpenCL specification is the result of industry collaboration in this area. Although adoption of this kind of computing has been swift in some sectors, the jury is still out as to
whether this method will become a mainstream tool in HPC,
but rest assured that companies such as AMD are working
hard to make it so.
The communication: Interconnects

In order to keep all those nodes busy, a good interconnect
(a port that attaches one device to another) is needed. As
always, it all depends on your application set. In general, most
high performance cluster systems use a dedicated and fast
interconnect.
High performance interconnects are usually rated by latency,
the fastest time in which a single byte can be sent (measured
in nanoseconds or microseconds), and bandwidth, the maximum data rate (measured in Megabytes or Gigabytes per
second). There are other numbers that help as well, like the
N/2 packet size. This number is the size of the packet that
reaches half the single direction bandwidth of the interconnect (a measure of how fast the throughput curve rises). The
smaller the number, the more bandwidth (speed) that small
24

packets will achieve. A final number to look at is the messaging
rate. This tells you how many messages per second an interconnect can send and is important for multi-core nodes because
many cores must share the interconnect.
Although the numbers listed are good for sizing up an interconnect, the ultimate test is your application(s). In most cases,
your application will communicate using MPI (Message Passing
Interface) libraries. This software is a communications layer on
top of the hardware. MPI implementations vary and the true test
is to run a few benchmarks. (For more on MPI, see the section
HPC glue: Cluster software.)
Fast interconnects also require switching so that any node can
talk to any other node. One way in which high-end switches are
rated, in addition to the point-to-point numbers mentioned previously, is bi-section bandwidth. This rates how well the switch
supports many simultaneous conversations. A good bi-sectional
bandwidth will allow all the nodes to be communicating at the
same time. Although not a necessity for all clusters, large parallel applications usually require a good bi-section bandwidth in
order to keep the processors busy. Keep in mind that large clusters use switch hierarchies that are composed of many smaller
switching units. (This can be an important factor when sizing the
number of nodes; extra layers of switching increase costs and
network latency because traffic must traverse more hops to
get from point A to point B).
In terms of available technology, high performance computing
networks are generally being deployed using two major technologies; InfiniBand (IB) and 10 Gigabit Ethernet. Of course,
if your applications dont require a large amount of node-tonode communication, standard Gigabit Ethernet (GigE) is a
good solution. GigE is often on the motherboard and there are
very high density/performance GigE switches available.
InfiniBand has been successfully deployed in both large and
small clusters using blades and 1U servers. It has a freely
available industry backed software stack (www.openfabrics.
org) and is considered by many to be among the best interconnects for clustering due to its low latency and high throughput. Another key feature of InfiniBand is the availability of
large multi-port switches like the Sun Data Center Switch 3456,
which provides low latency. A large port density allows for less
cabling and fewer sub-switches when creating large clusters.
25
What the flip are FLOPS?

The term FLOPS is used quite a bit
in HPC. Like any other industry, they
have their share of acronyms. FLOPS
stands for floating point operations
per second. A floating point operation is a mathematical operation
(addition, multiplication, square root,
and so on) on a number with a decimal point and exponent (for example,
1.2345 x 103). To put these numbers in
perspective, a simple desktop calculator will deliver 10 FLOPS and a typical desktop workstation computer
can achieve some where between
20 and 30 GFLOPS (1 x109 FLOPS).
The fastest computers in the world
have recently broken the 1 PFLOPS
(1 x 1015 FLOPS) barrier.
Though not always stated, FLOPS

can have different levels of precision
and can vary depending on the program used to measure them. In general, there are two types of precision
(single and double) used in HPC. The
difference is due to the amount of
memory used to store the numbers.
Obviously, a double precision floating point number has more digits
than a single precision number (for
those curious readers, single precision is 32-bits and double precision is
64-bits). As a result, double precision
calculations take longer than single
precision. Keep this in mind when
you see someone potentially overselling a FLOPS rating.
The interconnect bottom line is as follows. Understanding

application communication requirements allows you to optimize
your price-to-performance. That is, your budget will be a
balance of nodes to network numbers. A fast (and expensive)
network means you may have to buy fewer nodes and could
have less overall computing power. A slower (and less
expensive) network means you can buy more nodes and
acquire more computing power. The balance between these
two sides must be based on your application needs.
Remembering the storage

Storage is often the forgotten sibling of the HPC cluster.
Almost all clusters require some form of high performance
storage. The simplest method is to use the head node as an
NFS server. Even this simple solution requires that the head
node have some kind of RAID sub-system. Often, NFS doesnt
scale to large numbers of nodes. For this reason, there are
alternative storage designs available for clusters. Most of
these designs are based on a parallel file system that in turn
26

defines the type of storage hardware that will be required. A
good example is the Lustre parallel file system. Lustre is a tested
open-source solution that provides scalable I/O to clusters.
HPC applications notoriously create large amounts of data, and
ensuring that an archiving system is available is crucial to many
data centers. Moving data to tried-and-true backup technologies
(magnetic tape) is an important step in protecting your HPC
investment. A good archiving system will automatically move
data from one storage device to another based upon policies set
by the user.
Another area to watch closely is Flash. Flash modules can
be integrated directly onto motherboards, as is the case
with the Sun Blade X6240, or integrated using the PCIe bus.
Keep in mind that a poor storage sub-system can slow a
cluster down as much, if not more, than a slow interconnect.
Underestimating your storage needs and assuming NFS will
just handle it may be a mistake.
A final aspect of the storage hardware is making sure it
will work and integrate with the other cluster hardware.
Sometimes storage sub-systems are purchased from separate
vendors, causing incompatibilities that can often be traced
to software driver issues or even physical incompatibilities.
Many users avoid these issues by utilizing a top-tier vendor
with well designed and tested sub-systems for both their
compute and storage needs. The key here is to manage the
data throughout its entire life cycle, from the cluster to home
directories to bulk storage to tape.
The HPC optimized appliance

Dents and dings can cost. As any
manufacturer knows, dented products dont sell at full price, if at all.
Preventing product damage has
a direct effect on the bottom line.
Consider a leading manufacturer of
home appliances. Dents and dings,
no matter how small, can render
an expensive washing machine
unsellable. When the company
noticed a high percentage of its

washing machines were being
dented between the factory and
the retailer they turned to HPC for
some help. By simulating the transit
process, they not only identified the
problem, but redesigned the packaging materials and the clamps that
were used as part of the packaging.
Fewer dents, more money.
27
Racking and stacking

Compute nodes can take the form of 1U servers or blade
systems (for more on this, see Chapter 2). The choice is
largely one of convenience. Blade systems often have a slightly
higher acquisition cost but can offer much more manageability,
better density, and some power and cooling redundancy.
Regardless of which form of computer node you choose, you
should ensure that all cluster equipment is rack mountable
and can be placed in standard 19-inch equipment racks. The
standard rack chassis can hold 42 U of equipment (1U equals
1.75 inches or 44.45 millimeters). It is very useful to map out
your rack chassis space for all equipment and allow some extra
for future enhancements.
Power and cooling

Power and cooling has gone from an overlooked expense
to a critical factor in cluster procurements. A general rule
of thumb for forecasting power and cooling cost is that the
yearly cost to keep a cluster powered and cooled will equal
roughly one third of the purchase price of the cluster itself.
For this reason, looking for a green solution makes both environmental and economic sense. For instance, choosing a cluster that uses 45nm Quad-Core AMD Opteron HE processors,
each with 55-watt ACP (Average CPU Power) rating can be a
smart move. Factoring in up to 20 percent lower idle power
compared to similarly configured competing systems may just
make it a brilliant move.
Finding Cluster Software

To get your HPC system up and running, you need something
to run on it. Thats where software comes in. Linux is by far
the most common operating system that HPC users choose
and any other software needs to work well with Linux. Other
options include Solaris, which is open and has full support
for Linux binary compatibility, and Microsoft HPC Server.
28
Operating systems
The predominant operating system for HPC can be summed
up in one word: Linux. Prior to the advent of Linux, the HPC
or supercomputing market used UNIX exclusively. Linux
represents a plug-and-play alternative and doesnt add any
licensing fees for the compute nodes (which can be quite
large in number). In addition to the Linux kernel, much of the
important supporting software has been developed as part of
the GNU project.
The GNU/Linux core software is open-source (see Chapter 5)
and can be freely copied and used by anyone. There are,
however, requirements to ensure source code is shared. The
openness and shareability of GNU/Linux has made it an
ideal HPC operating system. It has allowed HPC developers
to create applications, build drivers, and make changes that
would normally not be possible with closed source.
Virtually all Linux installations are done with a commercial
or freely available software distribution package. While the
commercial availability of free software may seem puzzling,
most commercial open source vendors use a support-based
model. Youre free to look at and alter the source code, but if
you need support, you have to open your wallet.
Users may recognize some of the Commercial GNU/Linux
distributions such as Red Hat, SUSE, and others. There are
community versions (no support options) available as well.
Red Hat Fedora, Open SUSE, and CentOS are examples of this
approach. It should be noted that although these distributions are highly polished in their own right, they dont contain
all the software needed to support an HPC cluster. Cluster
distributions are available that fill this gap.
HPC glue: Cluster software

There are several types of software tasks that are needed to run
a successful cluster. These tasks include administration, programming, debugging, job scheduling, and provisioning of nodes.
29
No free lunch
In order to use the combined power of
an HPC cluster, your software needs
to be made parallel. By the way, the
same goes for multi-core as well. A
typical computer program is written
to run on a single CPU or core. It will
not automatically use extra cores or
nodes in the cluster because there is
no free lunch in HPC. To run in parallel, one must change the internal
workings of the program. There are
several ways to accomplish this task.
If you want to run your program on
multiple cores, then using pthreads
or OpenMP is a good solution. If,
on the other hand, you want to run
across multiple cluster nodes (and

the multiple cores in the node) then
using the MPI (Message Passing
Interface) may be appropriate. In yet
other cases, a blend of both may be
the right solution. Regardless, the
level of effort required depends on
the program.
Many commercial ISV (Independent
Software Vendor) codes are already
parallel and will work out of the box
on clusters. The same is true for
many of the open source applications, although youre often responsible for building the open source
applications yourself.
From a users perspective, programming is perhaps the most

important aspect of the cluster. The most important HPC tool
for programming is probably MPI (Message Passing Interface).
MPI allows programs to talk to one another over cluster networks (this allows individual nodes to coordinate their participation in the overarching task at hand). Without this software,
creating parallel programs would be, and was in the past, a
very custom (and likely time-consuming) process. Today, there
are both open and commercial MPI versions. The two most
popular open MPIs are MPICH2 from Argonne Lab and the
Open MPI project.
In addition to MPI, programmers need compilers, debuggers,
and profilers. The GNU software includes very good compilers
and other programming tools; however, many users prefer to
use professional compiler/debugger/profiler packages such as
those offered by Sun Microsystems (Sun Studio 12 for Linux),
the Portland Group (PGI), and Intel. All vendors supply their
own tools and cluster stack software.
30

Although users like to run programs, administrators are
required to keep the cluster lights on as it were. There are
tools to help manage cluster installation (See the section
Provisioning: Creating the cluster) and monitor cluster
operation.
File systems
Almost all clusters use the standard NFS file system to share
information across nodes. This is a good solution; however,
NFS wasnt designed for parallel file access (for instance,
multiple processes reading and wiring to the same file). This
limitation had become a bottleneck for HPC systems. For this
reason, parallel file systems were developed.
One of the areas where the open GNU/Linux approach has
served the HPC community is with file systems. There are a
multitude of choices, all of which depend on your application
demands. HPC file systems are often called parallel file systems because they allow for aggregate (multi-node) input and
output. Instead of centralizing all storage on a single device,
parallel file systems spread out the load across multiple separate
storage devices. Parallel file systems are very often designed
because they must be matched to a particular cluster.
One popular and freely available parallel file system is Lustre
from Sun Microsystems. Lustre is a vetted, high-performance
parallel file system. Other options include PVFS2, which is
designed to work with MPI. Cluster file systems cover a large
area. In addition to massive amounts of input, scratch, and
checkpoint data, most HPC applications produce large amounts
of output data that are later visualized on specialized systems.
One thing to keep your eye on is pNFS (NFS Version 4.1),
which is designed for parallel NFS access. Most of the existing
parallel file systems plan to support the pNFS specification
and bring some standardization to the parallel file system
arena. ZFS, a file system designed by Sun, offers some exciting
possibilities for HPC because it is the first 128 bit file system
with many advanced features (for all intents and purposes 128
31
bits means it will never hit any storage size limits). ZFS is the
fix for silent data corruption.
Sharing is caring: HPC

resource schedulers
Clusters usually have lots of cores and lots of users. Sharing
all these resources is no trivial matter. Fortunately the allocation of cores is done by scheduling software and not users
(thus avoiding the ensuing chaos that could occur). Depending
on the application, a scheduler may closely pack the cores
(for instance, keeping them all together on a small number of
nodes) or distribute them randomly across the cluster.
A cluster schedule works as follows. Like the old mainframe
days, all users must submit their jobs to a work queue. As
part of the submission processes, the user must specify the
resources for the job (for instance, how many cores, how
much memory, how much time, and so on). The resource
scheduler then determines, based on site-wide policies,
whose job gets to run next. Of course, it depends on when
the resources become available and thus, as users often find
out, being first in line doesnt necessarily mean being first to
execute. The resource scheduler is a critical part of the cluster because it would be almost impossible to share resources
without some form of load balancing tool. One important
feature of the scheduling layer enables administrators to take
resource nodes off-line for repair or upgrade and users are
none the wiser users rarely have a say in which nodes
theyre assigned. Additionally, if a node fails, the running jobs
that use that node may fail, but other nodes keep working and
the scheduler will work around the failed node.
There are several popular and freely available resource
schedulers. One popular choice is Sun Grid Engine from Sun
Microsystems. Others include Torque, Lava, and Maui. In the
commercial sector there are fully supported versions of Sun
Grid Engine, Moab, Univa UD UniCluster, and Platform LSF.
32
HPC crash test dummies

These brave devices endure real
crashes to help make cars safer for
the rest of us. Everyone agrees that
making cars safer can save lives,
reduce medical costs, and in general
make for a better world. The only problem is that crash test dummies, for all
their sacrifices, dont really tell us how
a crash will affect the human body. To
learn this, we will need to start testing humans instead of crash dummies.
Of course, Im talking about virtual
humans in virtual cars crashing in to
virtual walls. This idea is the next step

in automotive simulation. Using HPC
methods it will someday be possible
to determine the effects on all aspects
(for instance bones, organs, muscles)
and forms (for instance, adult, child,
male, female) of the human body. And,
the results would not stop at the virtual crash wall. A virtual HPC human
would allow a new epoch of design
and safety engineering.
Ready to run application software

There are plenty of cluster aware software packages in both
commercial and open source form. Many of the commercial
packages may require specific GNU/Linux distributions, but
open source packages can be built within your specific environment. Commercial packages may or may not have better
performance or feature sets; however, they all have professional support options that are not always available with open
source packages. Due diligence will pay off in this area.
Provisioning: Creating the cluster

Turning raw hardware into a functioning cluster is not as difficult as it was in the past. This process is called provisioning
and it involves installing and configuring the head node and
worker nodes of the cluster. There are two main methods that
are used:
Local: This type of provision involves placing a full
operating system image on each node. Obviously, this
method requires that each node have a hard disk drive
(or other persistent local storage). Care must be taken
to keep the node images synchronized (which consumes
valuable time and resources). The advantages to this
33
method are that the nodes can be started independently

and they dont require a central server; there can also
be advantages in quicker boot times and reduced peak
network traffic as a large number of nodes are brought
on-line.
Remote: This method configures each node across the
network when it boots using standard protocols such as
DHCP (Dynamic Host Configuration Protocol) and TFTP
(Trivial File Transfer Protocol). These methods often are
called diskless because they dont require a hard drive
on each node; however, they dont preclude using node
hard drives for local and temporary storage. The advantage of remote provisioning is that the software environment for each node is centralized and a single change
can be easily propagated throughout the cluster. In addition, in cases where nodes dont have hard drives, this is
the only way to bring up the nodes.
Cluster Tool Kits

Most GNU/Linux distributions lack many of the key cluster
software packages. To remedy this situation, there are now
cluster distributions that provide turn-key provisioning and
ready-to-run software. Many of these are also freely available
and based on some of the above mentioned distributions. The
freely available options include
Sun HPC Software, Linux Edition: www.sun.com/
software/products/hpcsoftware
Rocks Clusters: www.rocksclusters.org
Oscar (Open Source Clusters Application Resources)
http://oscar.openclustergroup.org
There are commercial versions as well. These are based on
GNU/Linux and often provide extra commercial applications
and support as part of the cluster package:
Scyld Clusterware
ClusterCorp Rocks+
Platform OCS
Red Hat HPC Solution
34
How to become a cluster expert

in one easy step
It is very simple to become a cluster
expert. The answer to every question
about clusters and HPC is basically
the same: It all depends. Clusters
allow so much user choice it really
does depend on your specific needs.
If anyone asks you a question about
HPC or cluster computing, simply rub
your chin and say It all depends.
From there, youre on you own. Start

by understanding your needs, ask
specific questions, and pay attention
to what others have done in your
situation (with todays infinitely connected world theres no shortage of
information at your disposal).
Chapter 4
Pulling It All Together

In This Chapter
Knowing where to start
Finding a trustworthy source
Following some guidelines
Looking at benchmarks
his chapter goes over some of the practical aspects of

putting together an HPC system for your company.
Following the HPC Recipe

So you want to purchase an HPC cluster. Where do you start?
The following guidelines may be helpful, because the HPC
procurement process is an exercise in confirming assumptions
and sorting out the details.
Start with the personnel

Do you have a staff that is well versed in HPC methods? In most
cases, you probably dont. It may be time to consider a consultant
or some educational alternatives. The Internet can be helpful here.
There is a plethora of information about HPC clusters on the Web
(see the HPC Web resources sidebar). You can also find mailing
lists and community forums where you can submit questions and
discuss issues. Chances are, someone may have been in a situation similar to yours. Depending on your area of interest, you
may want to ask your colleagues. At this point, you may also
want to contact some hardware vendors and ISVs (independent software vendors) that offer HPC clusters. These folks can
be helpful and may have pre-sales resources of their own to
offer. Be clear that you arent at the purchase stage just yet.
36
High Perfomance Computing For Dummies
Consider the software

Are you planning to use commercial applications? If so, you
probably want to contact the ISV if you havent already. They
will undoubtedly have previous customers that once were in
the same (or a similar) position as the one you currently face.
They are likely to also have some suggestions as to the type
of hardware and software that best supports their product
(with so many moving parts, compatibility can be important).
If youre using an open source solution, it may be time to
stop lurking on the mailing lists and forums. Instead of just
reading the forums, ask what type of design works best for a
given software application. Be prepared for opinions that may
differ. Good arguments are very informative!
In addition to the application software, also consider the software infrastructure. In many cases you will be using Linux. There
are a number of ways to provision a cluster (see Chapter 3) and
care should be taken to choose the one that fits your needs. A
few words of advice may be helpful here. It is relatively simple
to set up a cluster of Linux boxes with some clustering
software on the nodes. Success may come quickly, but in the
long term you may be creating a rats nest of dependency
problems. So careful planning at the outset can be important.
There are other issues that arise when deploying do-it-yourself
system software across clusters. First, how well does it scale?
Something that works for 8 servers may fail on 128 servers.
Second, consider how you will update and maintain the software across the entire cluster. Although these are basic system
administration issues, it is surprising how many administrators
trip over many of these issues after the cluster is up and running.
Specify the hardware

After you have a good feel for the type of hardware, start
specifying the exact hardware you will need. Dont be afraid
to ask a hardware vendor for some evaluation units to run
some benchmarks of your own. Verifying assumptions is very
critical before you spend any money on hardware. Depending
on the hardware vendor, they may have systems available
for external large-scale benchmarking (highly recommended
if possible). Next youll want to create a concise and specific
RFP (request for proposal); this is how you communicate your
needs to others interested in supporting you with your HPC
Chapter 4: Pulling It All Together
37
goals. Some companies can even help you asses your business needs. Sun Microsystems, for example, has Solutions
Centers worldwide to help with this. To find out more visit
www.sun.com/solutioncenters/index.jsp.
Who Is Going To Stand By Me?

In the early days of cluster computing, most systems were of
the do-it-yourself (DIY) variety. Although this method is still
popular today because of its relatively low cost, it may not be
the best way to achieve your HPC goals (especially if youre
just entering the realms of HPC). Another choice one that is
especially good for newcomers to HPC is to select a vendor
that will integrate and support the entire cluster.
HPC Web resources

Of course you can search the Web
with HPC keywords, but the following
sites will help you get started with
focused content and news:
AMD.com/hpc: Information about
AMDs HPC solutions designed to
help solve todays complex scientific and engineering problems.
Beowulf.org: Home of the original Beowulf project. It also hosts
the Beowulf mailing list, one of
the best resources for learning
and asking questions about cluster HPC. This is where the rocket
scientists hang out.
Blogs.sun.com/HPC: News, tips,
applications, podcasts, videos,
and more. A great source of HPC
news and information.
ClusterMonkey.org: An open
community-oriented Web site for
HPC cluster geeks (and non-geeks)
of all levels. There are tutorials,
projects, and many how-to articles

on the site. There is also a links
section with an up-to-date set of
links to essential cluster topics
including books, software distributions, and more.
HPCWire.com: A good site to
keep abreast of news, events, and
editorials from the HPC market.
Inside HPC: Another good site
for news about the HPC market.
LinuxMagazine.com: A general
Linux site with a large focus on
HPC. There are plenty of articles on
many aspects of HPC and Linux.
Scalability.org: A great blog that
provides real experiences and
opinions from the HPC trenches.
Sun.com/hpc: Information about
Suns HPC solutions, where you can
find info on compute systems, data
storage, and industry solutions.
38
Does it scale?
The term scalability is often used with
clusters (and parallel computing). It
basically means how many processors
you can throw at a program before it
will not go any faster. Some programs
have difficulty using even two processors while other use thousands. The
difference is scalability. Scalability
depends on the nature of the program
(the algorithm) and the underlying
hardware on which it runs. The situation is very analogous to framing a

house. Theres a particular number of
carpenters that will build the house
the fastest. Beyond that number,
there will be no improvement, and
possibly a slow down if you add more
workers. Remember Amdahls Law
(see Chapter 2 for more on this).
In todays market, a turn-key cluster has many advantages

and is well worth the additional cost (over the raw hardware).
A vendor that has access and expertise with a wide array of
hardware and software can be essential to successful deployment of an HPC system. For instance, a vendor that can offer
an integrated and tested parallel file system can be preferable
over one that must engineer it on the fly or buy it from another
vendor. Long term support can be essential as well. Over the
lifetime of your cluster you will likely have questions about
various aspects of the system. These may have to do with
upgrading, modifying, and expanding your cluster. A strong
partner will serve you well in this situation.
Another, often overlooked, area is local integration. Almost
all clusters are required to fit into an existing data processing
infrastructure. Making sure that the vendor has the capability
to assist with this process is also very important. Similar to
the last mile issue with Internet access, local integration can
require more work than anyone expected.
For these and other reasons, choosing a top-tier company
like Sun Microsystems as your HPC partner is highly recommended. Sun has a complete line of high performance hardware, including top-performance AMD processors that have
been qualified for HPC cluster use. Their past experience in
HPC and software contributions have helped move the HPC
market forward.
39
Fortran, really?
When most people ask about HPC
software theyre surprised to learn
that most HPC codes are written
in Fortran. Although many consider
Fortran to be an ancient language, it actually enjoys quite a bit
of use within the HPC community.
The reason for using the second
oldest computer language is largely
historical.
Many of the HPC programs were originally written in Fortran and users are
very reluctant to replace what works.
Indeed, some HPC programs are
composed of over one million lines
of source code, and some mature
codes have quite literally hundreds

or thousands of development years
invested in them! Converting all this
to another fashionable language,
besides being silly, is just not economically feasible. In addition, software tools, or compilers, understand
Fortran extremely well and are very
good at optimizing source code.
C and C++ are also popular HPC languages. Like Fortran, C and C++ are
close to the wires and can give the
user maximum performance in HPC
applications. (By the way, do you
know that Fortran stands for formula
translation?)
Following Some Dos and Donts

for HPC Cluster Procurement
Consider some of the best practices that can help ensure a
successful HPC procurement:
Put together a project plan. Include a plan that covers all
the aspects of the procurement including the people who
will touch the cluster (end users, administrators, faculty,
personnel, others). Pay attention to infrastructure costs
and find a home for your cluster before you buy it!
Kick the tires. If at all possible, run benchmarks and try
the hardware before purchasing anything. Most HPC
applications can be run on a few nodes. Some vendors
have test systems for running codes as well. Take full
advantage of this option. There can be bottlenecks, software incompatibilities, and other unseen issues that can
cause delays and add cost.
40
What is a Beowulf?
If you have heard the name Beowulf
mentioned in relation to HPC clusters,
you arent alone. However, what is surprising to some people is that a Beowulf
is not a cluster. It was the name of a
project at NASA where researcher Jim
Fisher accepted Tom Sterlings offer
to create a personal supercomputer.
Sterling and Don Becker then went

about creating a commodity cluster
supercomputer for under $50,000. The
project was a success and opened
the door to other Beowulf-style clusters
at many of the government labs. It
also helped establish Linux as the de
facto HPC cluster operating system.
Require acceptance testing. Create a set of criteria that

will ensure your cluster is working properly. Be sure to
include this in an RFP (request for proposal) you create.
It will give you peace of mind and eliminate the low-ball
hardware vendors that are looking for a quick buck.
Seek an HPC partner. Using a cluster vendor that has experience in deploying and supporting an HPC cluster can be
very beneficial, especially if you envision your relationship
as more of a partnership than a single transaction. Anyone
can sell you a rack server; few will ensure that it delivers a certain number of TFLOPS or will assist with issues
and questions after the hardware has been installed.
Furthermore, be wary of the low-cost hardware trap. The
industry is replete with disaster stories due to someone
trying to save a few dollars on memory or networking
components. When in doubt, remember that a quality HPC
vendor will know what works and what doesnt.
Ask for turn-key delivery. Although the vendor may
have to assemble a cluster at your site, make sure
that when they hand you the keys, it is ready to run.
Hopefully, you have included acceptance testing and you
will be up and running right away.
Here are few things that may hurt you in the long run:
Dont buy an HPC solution from a data sheet. Data
sheets can give lots of information, but a data sheet has
never calculated a single FLOP. Clusters are a lesson in
the details. There are always a large number of assumptions that can be verified by simply installing software
and running some basic tests.
41
HPC potato chips

Potato chips are about as low tech
as one can imagine. That is unless
youre a leading food processor who
manufactures potato chips. Because
the chips are manufactured and not
sliced, the company has control
over the shape and aerodynamics
of each chip. You may ask why the
shape of a potato chip is important.
Simple: Chips need to move at a high

rate of speed through the production
line. If they move too fast, the chips
fly off the line. Understanding how
the chips respond to rapid movement
means the manufacturing process
can be optimized. More chips in the
package and fewer chips on the floor
mean more profit.
Specifying untested hardware may not be the right

course of action. There is a tendency in HPC to want to
use the leading edge. A new motherboard, processor, or
interconnect may look attractive, and even show good
benchmark performance. However, knowing how well
new-to-market hardware will stand up to 24/7 operation
is always a gamble. The larger vendors always seem to
lag the smaller vendors with new hardware, but there is
a good reason for this difference. The top-tier vendors
will certify hardware and ensure it works as intended.
There is nothing that puts a lump in your throat faster
than racks of servers experiencing random failures and a
vendor that shrugs off your concerns.
Dont use single core benchmarks as a basis for processor or network comparisons. If youre going to run in
a multi-core environment then why not test in a multicore environment? Single core benchmark numbers are
basically meaningless if you want to run in a multi-core
world. Quite often, users are disappointed with real
world performance after they purchase the hardware
because they used inappropriate benchmarks in their
specifications. This difference becomes very apparent
when one considers the memory bandwidth of x86
processors.
Understanding Benchmarks
The purpose of a running a benchmark is to eliminate faulty
assumptions. Choosing a processor based solely on its SPEC
42

rating (see www.spec.org) or Top500 performance (see www.
top500.org) can be akin to a game of chance. For instance,
many HPC applications are sensitive to memory throughput.
Choosing a quad-core architecture with adequate memory
bandwidth may have a bigger influence on performance than
clock speed or SPEC rating. Benchmarks are useful indicators as
long as you understand whats being tested and the associated
results. Its important to understand your applications and workflow are the ultimate benchmarks for any cluster procurement.
In addition, not all multi-core processor architectures are
the same. Some are sensitive to process placement, which
is often out of the control of the end user. For instance, the
AMD Opteron processor has a symmetric quad-core design
coupled with a high performance and balanced memory architecture. This design helps achieve better scalability between
cores and across the cluster. Without proper benchmarking,
one might assume this is the case for all processor designs.
HPC coffee
Creating a good tasting coffee may be
art, but keeping it fresh is a science.
A leading food packager has found
that instead of metal cans, plastic
containers maintain coffee freshness longer after theyre opened.
The simple solution was to switch
to plastic coffee containers right?
Well not so fast. Plastic created
another problem. Coffee continues
to release gasses after packaging.
A metal container can easily sustain
the pressure, but a metal container
doesnt keep the coffee as fresh.
The solution was to design a check
valve that releases the pressure
that builds up in the plastic coffee
container. This solution, while fixing
the first problem, created a second
issue. When the coffee is shipped
to the stores, the truck can take

mountainous routes where the outside pressure becomes quite low.
Gases exit the container via the
check valve, but when the pressure
increases the container implodes.
To avoid crushed coffee the company applied HPC technologies to
engineer a container that would, for
lack of better words, implode gracefully. Using some clever design and
HPC, they were able to create canisters that can travel far and wide yet
still maintain their shape and keep
the coffee fresh after it is open. This
may be a good time to mention that
there is often lots of coffee-drinking
involved in the world of HPC too
good to the last FLOP.
Chapter 5
The HPC Community

In This Chapter
Understanding the open source model
Using open source
pen source software is used extensively with HPC clusters. In one sense openness has helped to foster the
growth of commodity HPC by lowering the cost of entry. It has
always been possible to cobble together some hardware and
use freely available cluster software to test the whole cluster
thing for little or no cost.
Understanding Open Source

Understanding free and open software is often a hurdle for many
organizations. The term free is often misunderstood and used
in the context of free as in free lunch (for instance, we dont
have to pay a license fee for this software). When most open
source practitioners speak of free they are talking about free as
in free speech. The distinction is important, and in the case of
HPC, a significant competitive advantage.
Free as in free speech software means the user has the right to
understand and modify the software as they see fit. In HPC this
provides three key advantages. First, software can be customized
in any number of ways allowing support for a multitude of file
systems, interconnects, peripherals, and other small market
projects that would otherwise never garner commercial interest.
Second, it provides a safeguard against unplanned obsolescence.
In the past, it was not uncommon for a large computer system to
lose software support because a vendor went out of business, the
system had been phased out, or the budget may have run out.
With open source, users have a choice to continue using these
systems and take over the support themselves. Finally, open
44

source allows a high level of optimization. Software installations
can be minimized so that only essential HPC software is present.
For instance, there is no need for a graphical user interface or
sound card drivers on cluster nodes. Customers like choice and
vendors like sharing the high cost of software development.
Openness also fosters community and collaboration. In the case of
HPC, there is a large ecosystem of users, vendors, and developers
from many companies and organizations that freely exchange ideas
and software without the need for legal agreements. This community also provides a huge and open knowledge base of ideas, best
practices, and experiences that help everyone involved. In HPC,
open just works better.
Sites such as http://developer.amd.com and www.
sunsource.net are examples of how AMD and Sun
Microsystems support open source.
Looking at How HPC Works

with Open Source
There are many types of open source licenses. For instance, the
GNU license from the Free Software Foundation has some requirements on how you must include the source code if you distribute
a program in binary form. Others may only require attribution
and recognition of copyright when binary codes are distributed.
In any case, the goal of open software is shared development and
use. In one sense, some projects, like creating and maintaining an
operating system, are such big jobs that it makes sense to share
the development effort across company and hardware lines. The
same can be said for the HPC market.
Although there are many arguments for and against open software, the HPC market represents an area where both open and
closed source solutions co-exist and work together. It is not
uncommon for an entire cluster to be composed of open source
software except for some commercial applications. For users that
compile their own codes, a commercial compiler, debugger, and
profiler are almost always present. Many clusters also employ
commercial schedulers that allocate resources on the cluster (or
groups of clusters) for an organization. Finally, though software
may be open and freely distributed there are a certain number of
companies that provide commercial support for cluster software.
Overall, the combined open and closed aspects of the HPC software ecosystem seem to work rather well.
Chapter 6
Ten (Okay, Six) HPC

Takeaways
In This Chapter
Understanding HPCs advantages
Developing a plan
oncluding this very brief introduction to the world of HPC

computing is rather difficult. Many topics were given only
cursory coverage, while others were totally omitted. It is possible however, to offer the following six HPC takeaways:
HPC can be a source of new competitive advantage: HPC
represents a new competitive advantage for many companies and industry sectors. Many organizations, maybe
even your competitors, are using HPC as a way to help
reduce cost, design new products and processes, solve
problems, and increase profitability. This trend is not
pie-in-the-sky hyperbole; it is good old fashioned R&D
with some powerful new tools.
Choices invite optimum design: Due to the growth of
commodity hardware, HPC has moved from a pricey boutique market to a cost-effective technology that has moved
into virtually every market sector. Many of the design
choices that were once made by a supercomputer vendor
are now made by the end-users themselves. In essence,
customer choice allows HPC solutions to be engineered for
specific problem sets and requirements, offering potentially
tremendous value and ROI for the end-user.
Try before you buy: Clustered HPC systems can be built
from commodity off-the-shelf hardware and software.
The cost to try out various configurations can be low,
46

but the payoff of making the right HPC decisions can be
immeasurably high. Benchmarking your application(s)
on actual HPC clusters is the only real test of true performance. Dont give too much weight to glossy data sheets
and industry-wide benchmarks or contests. To learn
how you can accelerate your computational engineering programs with your own HPC design and simulation
system, explore resources such as Clemson Universitys
Computational Center for Mobility Systems. This
resource allows you to see solutions in action and learn
more about the unique models and methods developed
there. See www.sun.com/solutions/hpc/eseminar.
jsp for more.
Develop a plan: Purchasing HPC resources isnt as hard
as it seems, if you follow some basic guidelines and ask
the right questions. There are both hardware and software vendors that understand your needs and will work
with you as a partner and not a parts supplier.
Join the community: Open source software and the
Linux operating systems have been big enablers for HPC
clusters. There is tremendous value in using communityderived software and expertise to help meet your project
goals. You may be pleasantly surprised to learn that many
top tier HPC companies are already active participants in
the community.
Have fun: One final bit of advice. Have fun. HPC is, if
nothing else, a great way to push the envelope and try
new and exciting ideas. The cost to play in the HPC world
is at an all-time low and there are plenty of resources to
help you have a good time.

High Performan Ce Computing: Making Everythi NG Easier!

Uploaded by

Copyright:

Available Formats

High Performan Ce Computing: Making Everythi NG Easier!

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

High Performan Ce Computing: Making Everythi NG Easier!

Uploaded by

Copyright:

Available Formats

What are some advantages of using HPC?

What are some advantages of using HPC?

What are some design choices available when using HPC?

What are some design choices available when using HPC?

g Easier!

Sun and AMD Spec

Douglas Eadline, PhD

SUN AND AMD SPECIAL EDITION

by Douglas Eadline, PhD

Chapter 1: HPC: Its Not Just for Rocket

Chapter 2: Getting to HPC ................................. 9

Chapter 3: HPC Building Blocks ....................... 21

High Performance Computing For Dummies

Chapter 4: Pulling It All Together ..................... 35

Chapter 5: The HPC Community ....................... 43

Chapter 6: Ten (Okay, Six) HPC Takeaways ...... 45

he quest to capture and manipulate the world in digital form

About This Book

High Performance Computing For Dummies

Icons Used in This Book

HPC: Its Not Just

ention the word supercomputer to someone and they

High Performance Computing For Dummies

HPC mud flaps

able to determine that trimming and

To many organizations, HPC is now considered an essential

The Forms of HPC

of thousands of servers all working together on a single

may be HPC systems out there that

High Performance Computing For Dummies

Who Uses HPC Today?

2007 HPC Market Share as

Supercomputers (Over $500K)

Successful HPC applications span many industrial, government,

Computer aided engineering (CAE): Automotive design

Cultivating the HPC edge

The CoC is a nonpartisan, nongovernmental organization based in

High Performance Computing For Dummies

Who Should Be Using HPC?

Of coffee, chips, and soap

can boast of significant cost savings

n the past when you wanted to deploy a supercomputer,

Enter the Commodity Cluster

High Performance Computing For Dummies

Chapter 2: Getting to HPC

essentially a plug-and-play replacement for UNIX the then

Its About Choice

High Performance Computing For Dummies

What Does a Cluster Look Like?

Chapter 2: Getting to HPC

High Performance Computing For Dummies

Gigabit Ethernet Switch

There are a few variations to be considered with any cluster

Chapter 2: Getting to HPC

Figure 2-2: Ranger provides computational capabilities to the national

Currently, there is a trend to move away from individual 1U

High Performance Computing For Dummies

If You Need Speed

Chapter 2: Getting to HPC