Você está na página 1de 21

THE DESIGN ENGINEERS

HIGH-PERFORMANCE
COMPUTING
HANDBOOK
OPTIMIZE YOUR PRODUCT DESIGN CYCLE.

Practical computing advice for design engineering teams.

P U B L I S H E D I N PA R T N E R S H I P W I T H
TABLE OF CONTENTS
PROLOGUE 4

CHAPTER 1 7
Create a Computing Workflow to Support Simulation-led Design

CHAPTER 2 22
Building a Balanced Workstation

CHAPTER 3 34
Parallelization Primer

HPC HANDBOOK

2 HPC Handbook | Table of Contents Table of Contents | HPC Handbook 3


PROLOGUE My real objective was to get the idea across: We have a technology that is going to
make electronics cheap, says Moore in this video, recalling the impetus of his predic-
tion.

What is the HPC Handbook?


The Design Engineers High-Performance ComputingHandbookis a multimedia resource
produced byDesktop Engineeringmagazinethat willshow engineering teams how
todetermine thebest combination ofcomputing hardware and engineering software for
their needs. The HPC Handbook is a living resource that consists ofcase studies, videos,
The number of transistors surveys, white papers and webinars hosted on hpc.deskeng.com;articles in
Desktop Engineeringmagazine; and regularly updated chapters in anexpanding electronic
incorporated in a chip will handbook. Chapters will include topics such as:
How to incorporate workstations, in-house clusters/servers and cloud
approximately double computing into one efficient engineering workflow.
How to balance workstation components and software preferences to get the

every 24 months. most bang for your buck when running specific engineering applications.
How parallel software paired with modern processors can improve engineering
DOWNLOADING
ONE CHAPTER GIVES
USERS ACCESS TO
FUTURE CHAPTERS,
productivity.
GORDON MOORE, Intel co-founder ALL AT NO COST.
What virtualization means to the design engineer and why IT wants to deploy it.
An analysis ofprofessional workstations vs. consumer PCs.
How to tap into additional computing power when on the go.
A roundup of the latest workstation hardware.

A
Simulation success storiesthat show how HPC enables simulation-led design in
s the world marks the 50th anniversary of Moores Law, its difficult not various industries.
to be amazed and even bewildered by the advances in computing tech- Simulation benchmarking studies.
nology. Today we take for granted thecomputing power now available in
cheap, easy-to-use smartphones. Soon, the power of supercomputers will
be just as accessible. What will that mean for design engineering? Why the HPC Handbook?
Design engineers are facing a more disruptive technological landscape today thanwhen
While most organizations dont have access to 36,000 cores today, it wont be long Moores Law was coined. The exponential increase in accessible computing powercou-
before these extreme core counts are commonplace, said Wim Slagter, product manag- pled with ubiquitous connectivity has fueled an exponential increase in product com-
er, High-Performance Computing (HPC), ANSYS in theFebruary 2015 issue ofDesktop plexity. Its no longer enough to design and develop products. Todays engineering teams
Engineering. And even todays users who are running at much lower core counts will are designing and developing systems from self-drivingcars to lightweight aircraft to
see direct benefits through considerably greater efficiencies. The results will be more all the connected devices that create the Internet of Things.
amazing products delivered to customers much faster than ever.
Advances insimulation-led design, optimization technologies and data management are
Achieving those results will require more than just advances in computing hardware. It helping design engineering teams innovate more quickly and powerful, affordable,
requires accessibility and affordability. Accessibility is enabled byengineeringsoftware accessible engineeringcomputing hardware and software makes it all possible.Thats
thattakes advantage of those hardware advanceswhilebeing so easy to use that theyre- whyThe Design Engineers High-Performance ComputingHandbookwas created. Design
quire no special knowledge to do so. Affordability is a relative term based on total cost engineering teams need a constantly evolving reference resource that explains
of ownership vs. return on investment, but theres no disputing the fact that computing whatworkstations, clusters and cloud computing can do when used with the latest
prices continue to decreaseas computing power increases. simulation, visualization and rendering software.

In fact, thats the pointMoore was trying to make when he inadvertentlylaid down the
law via a wild extrapolation.

4 HPC Handbook | Prologue Prologue | HPC Handbook 5


CHAPTER 1:
CREATE A COMPUTING
WORKFLOW TO SUPPORT
SIMULATION-LED DESIGN
Determine the right computing options
for every stage of design engineering.

6 HPC Handbook | Chapter 1 Chapter 1 | HPC Handbook 7


Y
ou are surrounded by computing power and functionality, so much so millions of lines of code. As the name implies, simulation-led design brings simulation
that it can be daunting to determine what type of resources to apply further forward in the design cycle to both reduce an organizations reliance on ex-
to different design engineering tasks. Are tablets, with their modest pensive and time-consuming physical testing, and to quickly try many different design
computing capabilities, useful for anyone on your design engineering approaches to find the optimal ones to pursue.
team? How should workstations be equipped to best support
conceptual designers vs. detailed computer-aided design (CAD) A simulation-led design strategy requires the right computing environment to support
software users vs. simulation analysts vs. rendering and animation experts? Should you it. As pressure mounts to be faster than the competition, it might be tempting to clamor
set up a cluster of computing resources? Should you burst to the cloud? What about for an immediate computing hardware investment. Those who hold the purse strings
virtual machines? The list of computing questions goes on and on. There are answers to in your organization may even agree. But before long, your engineering workflow may
be found, and theyre worth uncovering. include a patchwork of expensive computing solutions that is a nightmare to maintain, is
underutilized, and/or wont scale to support future engineering challenges. On the other
KEY WORDS Using the right tool for the job not only avoids frustration, it removes precious time end of the spectrum, those who hold the purse strings may not understand why the
cluster: A group of from the design cycle, ultimately resulting in faster product development. While every engineering teams computing needs should stray so far afield from the needs of other
servers built that design engineering team may have different computing requirements, they all share the departments. They may be reluctant to investigate, purchase and maintain high-perfor-
are connected
via certain same need you do: To develop better products faster in an increasingly complex product mance computing solutions that are critical to simulation-led design.
interconnect development environment. According to Roland Berger Strategy Consultants, product
technologies. complexity doubled between 1997 and 2012 across all industries, even as the average The solution is a balanced approach that looks at the big picture, puts the engineering
product lifecycle shortened by 24%.1 departments importance to the company in context, and wisely invests in the comput-
the cloud: A ing needed to support an efficient engineering computing workflow. Such a workflow
computing model Increasingly, engineering departments are meeting the challenge to produce better when approached systematically rather than in fits and starts, or as a template that
that enables users products in less time via simulation-led design workflows. For example, GMs PLM sacrifices engineering computing on the altar of enterprise conformity multiplies
to access and Leader Craig Brown said2 simulation and virtual testing has allowed the auto industry
configure to your organizations return on investment (ROI).
various degrees reduce its time to market despite the amazing complexities that have been introduced
a networked to automotive design from lightweight materials to more efficient engines to tens of
pool of computing
A systematic approach requires you to:
resources. 1. Take an inventory of needs.
2. Match those needs to the best computing resources.
3. Deploy those resources with proper user training.

STEP 1: Get What You Need

T
he process of taking an inventory of needs will vary from organization to
organization. A small engineering team may get everyone in the same room to
get the groups needs out in the open. A large, engineering-centered enterprise
may opt to create a planning team or hire a consultant to observe, survey and interview DOWNLOADING
its engineers to reveal the same information. No matter what form the process takes, its ONE CHAPTER GIVES
USERS ACCESS TO
important to differentiate between wants and needs. FUTURE CHAPTERS,
ALL AT NO COST.
Needs are simply the differences between your current achievements and your desired
accomplishments, according to A Guide to Assessing Needs3, a free resource available
via theWorld Banks open knowledge initiative that deeply delves into the process of

3 Watkins, Ryan; West Meiers, Maurya; Visser, Yusra Laila. 2012.A Guide to Assessing
1 Mastering Product Complexity, Roland Berger Strategy Consultants. Needs : Essential Tools for Collecting Information, Making Decisions, and Achieving Devel-
2 2015 Siemens PLM Connection Americas User Conference presentation. opment Results. World Bank. World Bank. https://openknowledge.worldbank.org/han-
dle/10986/2231 License: CC BY 3.0 IGO.

8 HPC Handbook | Chapter 1 Chapter 1 | HPC Handbook 9


needs assessment. Needs do not, however, include any mention or discussion of com- GPU CONSIDERATIONS
puters, budgets, training courses executive coaching, leadership, incentives, policy
analysis, microfinance strategies, holiday bonuses, reengineering, or any other tech- Some engineering workflows Quadro graphics card and an NVIDIA
niques used to achieve results. Rather, your needs are the basic gaps between current depend on maximizing Tesla companion processor, which NVIDIA also offers a means to
and desired performance. workstation performance without in turn frees up CPU resources remotely access its GPUs. The
relying on additional HPC resources. for additional work. The Quadro NVIDIA Grid Visual Computing
So this is not the time to say: I need a new workstation. This is, however, the time to In such a workflow, engineers family is specifically developed for Appliance is a GPU-based system
say: If were going to meet our deadlines, I need to complete simulation runs faster are tasked with carrying out graphics-intensive applications like that runs complex applications such
than the six hours they are averaging. simultaneous design and CAE CAD, while Tesla is designed for and sends their graphics output
simulation on a single workstation. parallel-processing jobs commonly over the network to be displayed on
A thorough needs assessment will likely uncover a number of gaps in your workflow Workstations equipped with required in simulation. Maximus a client computer, according to the
between where you are and where you should be. Some of those gaps can be closed NVIDIA Maximus technology can transparently and automatically company. The appliance literally
with new hires, training or process improvements. Others can be closed with the right turbocharge parallel processing. assigns visualization and simulation a box housing a set of GPUs can
technology. or rendering work to the right support anywhere from 8 to 16
NVIDIAs Maximus technology processor, enabling engineers to users.
enables a simultaneous design and simulate on the Tesla GPU without
STEP 2: Consider Computing Resources simulation workflow by distributing affecting their ability to design Kenneth Wong, Jamie Gooch

F
or those need gaps that can be closed with computing technology, its important tasks to two GPUs: an NVIDIA using the Quadro GPU.
to consider more than just hardware. Hardware is only made useful by software,
so the software youre using now, the software you should be using now, and the
software you plan to use in the future should help guide your hardware investments.

Conceptual Design easy geometry editing. If you want fit [a product that must connect, join, or fit into a
Conceptual design software applications have become much more than digital nap- predefined space], you may need a parametric modeler. If you want function [the opera-
kins, as theyre sometimes called. In Form, Fit or Function, which appeared in the tion of the product], you could explore simulation, he says. KEY WORDS
May 2015 issue of Desktop Engineering, Dr. Ken Versprille, executive consultant at parallel processing:
CIMdata, said conceptual design means different things to different people. While conceptual designs definition has broadened beyond sketching out and sharing dividing program
instructions
If your major criteria is form [the shape or the look of the product], you would want ideas, its still the main class of engineering software that makes use of tablet comput-
among multiple
ings mobility and touchscreen interfaces (see Digital Napkins sidebar). Tablets can fit processors in order
well into a design engineering workflow in situations where mobility is critical, such as to complete the
instructions faster.
capturing data in the field.

Specialized simulation apps a fairly recent development are also making tablets
more useful to design engineers. Such apps, often created by simulation analysts, allow
TABLETS non-experts to change variables and run particular simulations. The apps can be particu-
larly useful on the shop floor or in customers offices.4
+ PROS:
More portable than laptops
When deciding on a tablet or ultraportable, consider the fact that an ultraportable lap-
Touchscreen interface
top or tablet-laptop hybrid may fulfill both needs for a high level of mobility and basic
Long battery life
CAD functionality, albeit at a slightly higher price and weight than a tablet.
Many specialized apps

- CONS:
With the advent of Windows 8, most PC manufacturers released tablets and convert-
ible systems combining touchscreens with small, often detachable keyboards, wrote
Limited functionality
Limited processing power
Limited onboard storage
4 For more information on simulation apps, see Desktop Engineerings Expert-Built Apps
Limited expandability
Expand the Reach of Simulation, The Appification of Simulation, and Engineering for
Everyone.

10 HPC Handbook | Chapter 1 Chapter 1 | HPC Handbook 11


David Cohn in the December 2013 issue of Desktop Engineering5. One new category of
systems is the ultrabook. Based on a specification developed by Intel, ultrabooks rep- DIGITAL NAPKINS
resent a new class of high-end subnotebooks designed to feature reduced bulk without
compromising performance and battery life. When used on a tablet with into CorelCAD as model space
a stylus, apps like PTCs objects with a representation in
In 2014, Cedric Desbordes, Graeberts marketing executive, told Desktop Engineering6 Creo Sketch, Autodesk Sketch- a layout sheet.
that he believes engineers and designers wont work exclusively in mobile devices. Book and CorelDRAW closely
Instead, theyll work in a cross-device CAD experience, switching to different devices mimic the 2D pen-on-paper One sketching program that
as circumstances demand. For example, they may approve a change from their phone, workflow, coming closest to a bridges 2D and 3D is CATIA
cocktail napkin. PTC Creo Sketch Natural Sketch from Dassault
offers Spline tools in addition Systmes. The program lets you
5 Review: Mobile Tablet/Notebook Convertible Can Run CAD, David Cohn, Desktop Engineer- to painting tools; the product is use the traditional hand-sketch-
ing, Dec. 2013 meant to augment PTCs other 3D ing techniques, but on different
6 Mobile CAD Evolves, Kenneth Wong, Desktop Engineering, Nov. 2014
design apps, like PTC Creo Direct working planes in 3D. You may
or Creo Parametric. also use the familiar extrusion
and projection techniques in the
With digital brushes that repro- you can export Spline curves for sketching environment. With this
TABLET COMPARISON duce the look of oil, watercol- further development in Creo approach, you can transform a
or, pastel and other mediums, Parametric. With art-centric 2D rough sketch into a 3D model
OPERATING KILLER Autodesk SketchBook offers a SketchBook, you cannot export with volume and mass.
SCREENS WEIGHT SPECS COST BATTERY
SYSTEM FEATURES richer environment for artists Spline objects; therefore, the use
and can function independently of the sketches created in it may Recently, the last of the big four
iOS 7.9-12.9 inches; 0.68-1.57 Storage: Up to $250- About Largest App Store
of other 3D design programs. be limited to background layers engineering software compa-
1024x768, 2048x1536 lbs. 128GB $1,079 10 hours
Examples: iPad mini, or 2732x2048 Easiest synching Neither Creo Sketch nor Sketch- to guide your 3D modeling. nies, Siemens PLM Software,
iPad Air 2, IPad Pro resolution RAM: up to 4GB within Apple Book gives you the ability to announced it too would release
ecosystem draw with dimensions (the way The CorelDRAW Technical Suite a conceptual design package
Chip: Up to 64-
bit A9X, est. at Works with Apple you would typically do in 2D X7 includes CorelDRAW for named Catchbook that it prom-
2.25GHz Pencil ($99) CAD), but thats not necessarily vector illustration and page ises will let you capture ideas,
a drawback if youre mainly con- layout projects, Corel DESIGNER draw accurately and share your
cerned with rough ideas. You can for technical illustrations, and creations. Catchbook is expected
Android Up to 24 inches, but Varies Storage: Up to $150 - Up to Customization export the sketches as images the Corel Photo-Paint image to be released in the fall of 2015.
more commonly in 128GB $1,500 13 hours into CAD programs for further editor. CorelDRAW and Corel
Examples: Hardware 7-10 inch sizes Large choice in
from HP, Dell, Lenovo, RAM: Up to 3GB hardware vendors development. With Creo Sketch, DESIGNER files can be imported Kenneth Wong, Jamie Gooch
HTC, Samsung, Sony, Up to 2560x1600
LG and many others resolution Chip: Choices
include
NVIDIA Tegra
K1, and more gather measurements onsite with a mobile tablet, and edit the DWG file back in the
office on a workstation.
Windows About 10-12 inches, 1.37 lbs. Storage: Up to $500- About Full Microsoft
up to 3200x1800 1.76 lbs. 512GB $1,800 7-10 hours Office support Windows 10 promises to further integrate mobile computing into the workflow by pre-
Examples: Microsoft resolution senting a common user interface on workstations, laptops, tablets and phones. Whats
Surface Pro 3 RAM: Up to 8GB Surface Pen Stylus
more, the new operating system is expected to include features that make transferring
Hardware also Chip: Intel Atom Mouse Support data back and forth among devices more seamless.
available from Dell, to Intel Core i7
Lenovo, ASUS and Hardware like Microsofts Surface Pro 3, with its detachable keyboard and ability to
others
run the full version of Microsoft Office, is blurring the line between tablets and laptops.
Still, convertibles and ultrabooks are not intended for creating and working with the

12 HPC Handbook | Chapter 1 Chapter 1 | HPC Handbook 13


using both AutoCAD 2010 and AutoCAD 2015.

It took more than 10 hours to complete the five drawings using AutoCAD 2010 com-
pared to 6.5 hours to complete the same five drawings using AutoCAD 2015, represent-
ing time savings of 36%, without any change to the computer on which the software
WORKSTATIONS was run, according to the report, Getting the Most from AutoCAD 2015 with Dell
PROS: Precision Workstations. When the workstation was upgraded to the more modern Dell
+ Versatile Precision T1700, the time required to complete the five drawings using AutoCAD 2015
Affordable Power was further reduced to 5.3 hours, a total time saving of 48% compared to using Auto-
UPGRADING FROM OLDER
Ease of Setup and Use CAD 2010 on an older workstation. WORKSTATIONS OR FROM
CONSUMER PCs TO PRO
Expandable/Customizable WORKSTATIONS
Another example shows similarly impressive results. Lenovo pitted a 2010 ThinkStation PROVIDES A QUICK
- CONS: S20 against a 2013 ThinkStation S30 running SolidWorks 2013. Both workstations were ROI.

Software patches often equipped with graphics cards from the middle of the range of cards they support, sol-
consist of individual updates id-state drives (SSDs) and the same amount of RAM. The newer workstation completed
Not easily portable the SPECviewperf12 benchmarks for SolidWorks 2013 more than twice as fast as the
Not infinitely scalable three-year-old workstation.

Because of the workstations central place in most design engineers work, upgrading
consumer-level PCs to professional engineering workstations (see High-end PC vs.
larger models typical to many industries. In the engineering workflow, tablets are often Entry-level Workstation sidebar), or investing in new workstations every three years
confined to quick concept work and reviewing designs on the go while ultrabooks add can provide a big speed boost to support the latest versions of software for detailed
some additional CAD functionality. However, cloud computing, Software-as-a-Service CAD work as models get larger.
(SaaS) and specialized apps continue to evolve, making it possible to even do more with
mobile computing power even some simulation work. Unlike ultraportables and tablets, workstations are designed to be customized. Most
vendors offer purpose-built baseline models with the option to change the processor
and graphics, add more memory, use different hard drive technology and more. The
Detailed CAD challenge is to buy what your particular needs require without over investing in features
KEY WORDS
The bulk of CAD work is still performed on engineering workstations that come in you wont use.
cloud computing: small, desktop form factors as well as larger, deskside towers and mobile variations, each
Using remote
computing with a number of different internal configurations of processors, storage, memory and If your needs include anything beyond CAD and small simulation runs,
resources, such graphics. This versatility is what has made workstations the workhorses of engineering the optimal workstation recommendations will change. Small simula-
as a network of design. There are very few, if any, engineering workflows that would be improved by tions are those in which you set up models, add and couple physics, and HPC RESOURCES
servers via the
Internet rather than completely bypassing the workstation. investigate different scenarios. + PROS:
a local computing Scalability
resource.
If your inventory of needs includes speeding up design time, then investing in both the latest Cost saving potential
software and the latest workstations is an option worth investigating. Software vendors are Simulation and Rendering Flexibility
SaaS: Stands for constantly updating their engineering applications to take advantage of the latest hardware For most engineers, the question isnt whether to incorporate workstations Reduced software
Software-as-a- improvements. Investing in one without the other will not allow you to realize your full per- into a simulation-led design engineering workflow, but when to incorpo- maintenance
Service, which
formance potential. Using outdated hardware has even more significant consequences when rate other high-performance computing (HPC) options such as servers,
describes a means
of subscription working with larger, more complex models. clusters and cloud computing. Often, those needs coincide with the need - CONS:
software licensing to do more simulations simultaneously to arrive at an optimized design, to Can be daunting to setup
and delivering the Overcoming security
software from a
For example, Desktop Engineering Contributing Editor David Cohn worked with Au- simulate more complex models, or to create more realistic renders.
todesk and Dell to test the productivity gains a typical user would experience when concerns
central location.
upgrading to Autodesk AutoCAD 2015 software. Using older systems and a newer Dell Many of our customers come to us after they have lost projects totheir Dependent on connectivity
Precision T1700 workstation, Cohn timed the repeated creation of common drawings, competition on a time-to-market basis, Dominic Daninger, vice Often dependdent on
off-site support

14 HPC Handbook | Chapter 1 Chapter 1 | HPC Handbook 15


WORKSTATION RECOMMENDATIONS*
SIMPLE COMPLEX
HIGH-END PC VS. ENTRY-LEVEL WORKSTATIONS
CAD RENDERING
SIMULATION SIMULATION/CAE
Theres virtually no difference in price be- that involves hardware and software vendors work-
Multiple processors, tween entry-level workstations and con- ing together to find and solve any compatibility
4-core processor; high 4-core processor; high As many cores as
Processor 6 processor cores; high sumer PCs that seem comparable on paper. issue before the workstation is released.
clock speed clock speed you can afford
clock speed
So whats the difference?
Memory 8-16GB 16GB 16-32GB 16-32GB Independent Software Vendor (ISV) certification
A surgeon doesnt wake up in the morning and pick ensures the system is qualified and supported by
Storage SSD SSD SSD SSD up a kitchen knife and go to work to perform brain the ISV for workstation software applications, says
surgery, says Andy Rhodes, executive director of Al Makley, director of ThinkStation Architecture
Graphics Entry-level Mid-range Mid-range or higher Mid-range or higher Dell Precision workstations in Making the Case for and Technical Solutions, Lenovo. Desktop PCs do
* The caveat: The frequency with which you work with large, complex models; perform advanced simulations; and create photo- Professional Engineering Workstations. They use a not have ISV certification.
realistic renderings will largely determine the best workstation configuration. scalpel. The two things are both knives, but one is a
professional tool for a professional job. We certify workstation-class graphics cards as
part of our hardware certification program, says
president of Engineering, Nor-Tech, told Desktop Engineering7. The company manufac- Even high-end consumer systems are not ideal for Justin Kidder, director of SolidWorks Graphics and
tures and integrates workstations, servers, clusters and storage. They may not be using professional engineering work, according to Rich- Architecture. What this means is, we know what to
design simulation or if they are, they may be doing it on workstations and waiting 72 ard Runnells, director of Marketing for Solid Edge expect in terms of graphics drivers and their inter-
hours for each solver simulation solution. at Siemens PLM Software. actions with the workstation. While some custom-
ers may choose to run on consumer or gaming PCs,
Because modern workstations are so powerful, its possible to configure them to produc- While in some instances using a high-end we do not certify them for use with SolidWorks, and
tively handle all but the most demanding simulation runs and renders. However, config- consumer PC or even a gaming PC may work to do not recommend that our customers do this.
uring such a high-powered workstations can cost upwards of $12,000, and they may not meet the minimum requirements for desktop CAD
provide a decent return on investment especially if you only have an occasional need applications, typically these systems are maxed out Matthew Burke, principal and CEO, M2 Technolo-
for that much power. In that case, it often makes sense to burst to a public or private and provide little to no room for future expansion, gies, Autodesk Reseller, sums it up in the July 2015
HPC resource. he says. The cost of an entry-level workstation, issue of Desktop Engineering magazine: CAD is
which comes pre-configured and certified to run expensive, and designers and engineers are expen-
If the simulation software you are using is optimized via parallel programming, which CAD applications, such as Solid Edge, is in many sive give them a machine with the architecture
involves splitting operations into parts that execute simultaneously on different processors, cases lower in price than that of a top-of-the-line capable of allowing them to perform their jobs
you can benefit from more workstation processor cores, or as many cores and nodes as gaming PC. efficiently, he says. To put it another way, dont
you want on a cluster or the cloud. Most major CAE software vendors have products that ask Jeff Gordon to win at Daytona in an SS sedan
take advantage of HPC resources with multiple cores from CPUs and GPUs. HPs Jeff Wood, vice president of Product Manage- from the local Chevy dealer.
ment for Workstation & Thin Client Business at HP,
Such resources may take the form of a cluster (virtual or physical), data center or the agrees with Runnells conclusions. Kenneth Wong, Brian Albright
cloud (pubic or private). While the IT infrastructure, setup, maintenance and costs
for each are different, the workflow benefit is the same: Intensive computing tasks are Buyers are missing a huge benefit in productivity
removed from the workstation and carried out on the HPC resource. Because the HPC when they opt for the consumer PC, certainly ver-
resource can be configured with many times more computing capability than even a sus our Z workstations, he says. The workstations
top-of-the-line workstation, intensive computing tasks can run more quickly while have been highly tuned for professional applica-
allowing you to continue to use your workstation at peak efficiency. tions. Another consideration is [that] many of the
CAD software vendors are unwilling to provide
software support for non-certified consumer PCs.
Public Cloud Computing
Imagine a CFD study that takes four hours to run on your workstation, and you want Thats because certification is a rigorous process

7 Computer Cluster Considerations, Jess Lulka, April 2015, Desktop Engineering.

16 HPC Handbook | Chapter 1 Chapter 1 | HPC Handbook 17


to do 50 of them. Simple math tells me it is going to take five weeks or more do each of that need site-wide collaboration across the country and want common file storage, but
these, says M2 Technologies Burke. Now imagine setting up these 50 studies to run dont have the IT department to set that up. Its also good for overflow needs, if you
on the cloud and getting an e-mail in four hours that all 50 are available. That elastic need four or five extra seats for one week. Also, if you are running a big analysis job, but
computing capacity is the real power of the cloud. you dont want to buy a machine because you only need it for a week.

If you have modern workstations, but your simulation and/or rendering workload is still The shadow looming over cloud computing since its inception has been security, but
bogging them down and youre using software that is able to take advantage of mul- that concern is lessening. Security at popular public cloud computing platforms, such
tiple processing cores then offloading certain intensive computing tasks is a workflow as Amazon Web Services and Microsoft Azure, and service providers is generally better
slam dunk. However, those are a lot of criteria, and even if theyre met, which type of than what most SMBs (and even many large enterprises) can support in-house. Howev-
HPC resource should you turn to? er, some companies are still leery of putting their intellectual property in the cloud. For
those companies, a private cluster is a practical option.
As a general rule of thumb, public cloud computing is especially well suited for small-
and medium-sized businesses (SMBs). The business case is hard to resist: virtually
unlimited computing power available as needed without a significant IT infrastructure Clusters/Private Cloud Computing
or staffing investment. Intellectual property protection is one of the key reasons that customers choose to go
with an in-house private cluster vs. a cloud solution, Nor-Techs Daninger said. An-
Chris Teague, senior applications engineer at Saratech, a Solid Edge Reseller headquar- other reason is very few cloud HPC offerings have InfiniBand fabric (for networking).
tered in Mission Viejo, CA, outlines the perfect cloud computing customer in Desktop Cloud solutions typically will offer 10GB Ethernet, which has much longer latency KEY WORD
Engineerings July 2015 issue: Certainly the guy who is running part-time CAD and times than current InfiniBand fabric, he said. InfiniBand: A
cant justify buying the software and hardware all at once, he says. Smaller companies fast networking
communications
Rod Mach, principal at TotalCAE, a provider of private cloud and engineering IT standard used
services, said private clusters also allow companies to tap into HPC resources without within and between
modifying their engineering workflow. There are no extra CAE licensing complexities, high-performance
computers.
all work can be done on the same central system, he said. An on-premise professionally
KEEP MOBILE IN MIND managed cluster is much cheaper for your baseline constant engineering computation
needs.6
If youre not prepared to work anywhere, your At the same time, engineers appreciate the ability to
competitors might be. Mobile workstation take their work on the road, or even just down the hall. SMBs arent the only organizations weighing their HPC options. Even large enterprises
speeds increasingly compete with basic desktop work- With the prices for mobile power dropping, design with their own data centers should consider on-demand HPC resources for the engi-
stations. More and more, engineers are finding they engineers are finding fewer reasons to not make mobile neering department. Shared, enterprise-wide resources can easily be consumed with
dont have to choose between power and portability. part of their design arsenal. multiple simulation runs. For many companies, offloading the most intensive tasks to a
Jamie Gooch dedicated HPC resource whether onsite or in the cloud results in a net gain for
According to Tom Salomone, Engineering and AEC and productivity because smaller jobs can be run via internal servers and arent being held up
Manufacturing Segment Marketing Manager at Lenovo, by the larger ones.
more than two-thirds of design engineers with mobile
computers also use a desktop workstation. The key is to an optimal workflow is to find the hardware and software configuration
that is optimal for your specific environment, without purchasing more than you need,
Most people have mobile and desktop workstations, or investing in a solution that will not provide sufficient performance improvements to
Salomone said in a white paper produced by Desktop justify the additional cost. Take a balanced approach to get the most productivity for
Engineering8. Mobile still has some limitations. People your investment.
want bigger screens and access to more options; whats
happened is they are putting more on the desktop to
get better productivity, and you cant carry all
of that.

8 Workstations, Now is the Time to Upgrade


white paper, August 2014

18 HPC Handbook | Chapter 1 Chapter 1 | HPC Handbook 19


on-demand licenses weekly, monthly, depending on the vendor, says Joris Poort, CEO and
STEP 3: Deploy cofounder of San Francisco-based Rescale. Thats a big part of what we help solve for our

Y
ouve analyzed your needs, determined the right hardware for your software customers. We tell them not to worry about the licensing issues. We work with vendors
and decided on the best configuration for you computing resources. Youre Dassault, MSC, ANSYS, and others to figure out the best licensing plan for our clients.
done, right? Wrong. Your needs analysis should have determined why a partic-
ular computing resource is needed, but that doesnt mean your team will automatically If you like the idea of a turnkey solution, HPC Appliances are also an option. Physical
understand how to use those resources effectively, or how to navigate software vendors appliances provide private, in-house access to more computing power, while virtual
licensing requirements, which can quickly become complex. appliances do the same on the cloud. Both types of appliances are essentially pre-con-
figured hardware clusters designed specifically to plug and play with a particular design
Even putting something as simple an inexpensive computer tablet in the hands of the engineering software platform. HPC Appliances are an attractive option for simu-
engineers on your team is a waste of money and time without proper training. Likewise, lation-dependent businesses that dont want to invest in full-scale data centers, but
an engineer given an upgraded workstation may not fully realize and therefore not arent prepared to work exclusively with SaaS either. For those caught in the tug of war
MANY ENGINEERING TEAMS fully utilize its capabilities. Many engineering teams drop the ball when it comes to between in-house servers and the cloud, the plug-and-play simulation hardware might
DROP THE BALL WHEN IT just be their safe harbor.
COMES TO TRAINING AND training and support.
SUPPORT.
Finish strong in the deployment stage to make the most out of your local hardware.
Review
On the other end of the spectrum, many engineering teams are wary about deploying This chapter of The Design Engineers High-Performance Computing Handbook was writ-
cluster-based solutions. In the past, HPC clusters were notoriously difficult to imple- ten to provide a broad overview of the computing options available to support a sim-
ment, requiring specially trained IT personnel to set up, manage and configure jobs. ulation-led design workflow. Computing solutions geared toward conceptual design
engineering, CAD, CAE and rendering were provided to allow you to investigate the
In a typical HPC cluster scenario, a variety of open source tools are employed one options that will fulfill your specific needs.
program might handle provisioning of the cluster nodes while another package deals
with network configurations and yet another tool takes on scheduling and load balanc- Future chapters will delve more deeply into topics such as building a balanced worksta-
ing. Each package is equipped with a different user interface and corresponding learning tion, parallelization, virtualization, mobile computing and more. As a subscriber to the
curve, which means someone in IT has to be versed in the entire portfolio of tools or HPC Handbook, you will be notified when future chapters are available. In the mean-
a firm must have several HPC specialists on staff, which is cost prohibitive for most time, please visit hpc.deskeng.com for more high-performance computing resources for
design engineering teams.
SMBs. To add to the challenge, clusters have become increasingly complex as the num-
KEY WORDS ber of nodes and cores grows and the mix of components on which to distribute jobs
PaaS: Stands expands to include both CPUs and graphics accelerators.
for Platform-
as-a Service, a
computing model
Yet advances in HPC configurations and cluster management software are addressing
in which hardware these barriers to entry, making it far easier for smaller companies to get on board. For
and software tools example, IBM Technical Computing is now delivering high performance systems inte-
are provided to
users as a service. grated with workload and resource management tools that put HPC capabilities within
reach of SMBs from both a price/performance and ease of manageability perspective.98

HPC Applicances: Many service providers109 also specialize in helping engineering teams get up and run-
Hardware pre-
configured with ning on clusters or the cloud. Some even offer popular engineering design and simu-
software, often to lation software-as-a-service (SaaS) and Platform-as-a Service (PaaS) implementations
perform a specific that help customers, large and small, create and run customized, repeatable, multi-solver
function such as
providing additional simulation and optimization cycles.
compute power
for simulation or We can run [a simulation] using the customers existing licenses, or we can run it using
rendering.

9 For more information on clusters, see Desktop Engineerings The Future is Now white
paper.
10 See Desktop Engineerings Services Directory for providers that help setup HPC resources:
deskeng.com/services

20 HPC Handbook | Chapter 1 Chapter 1 | HPC Handbook 21


CHAPTER 2:
BUILDING A BALANCED
WORKSTATION
Choose the best hardware components
for the design engineering work you do.

22 HPC Handbook | Chapter 2 Chapter 2 | HPC Handbook 23


N
A modern high-end CAD workstation will have a multi-core CPU, Hyper-Threading and
obody likes a sluggish computer, but for designers and engineers,
a powerful GPU, along with a solid-state drive (SSD) instead of traditional hard drive to
a workstation that gets bogged down means valuable design
speed data access. But the minimum hardware requirements recommended by software
optimization time is ticking away. To keep engineers as productive
vendors dont always apply to every machine. Not every workstation necessarily has to
as possible, companies need to invest in professional-grade
have the best graphics. And just throwing a lot of money at the problem in the form of
workstations, but not every unit is created equal.
multiple processors with lots of cores doesnt guarantee performance improvements.
KEY WORDS
The annualized cost of a high-end workstation on a two- to three-year refresh cycle is Dont spend a lot on one component and sacrifice in other areas, says Scott Hamil- Hyper-Threading:
normally much lower than the annualized cost of a users CAD software, yet the correct ton, Precision specialist at Dell. In most cases, you get the most benefit from a system Intel Hyper-
hardware can have a profound impact on CAD user productivity, and the wrong hard- where all the components work well together. As long as you look at the workflow of Threading
Technology
ware can bring productivity to a halt, says Sean Young, worldwide segment manager, your users and the applications they are using, youll wind up with something good. enables multiple
product development and AEC at HP. independent
The types of applications you need to run will guide selection, as will determining what instructions to
run on each core.
Intels Andy Parma, Workstation Segment Manager, says engineers should consider apps will be running simultaneously. Photo rendering and simulation have a big impact One physical core
balance when considering a new workstation. When a workstation is balanced, on hardware configurations, for example, along with the size and complexity of models appears as two
processors to the
it provides excellent performance across four key elements: processor, and assemblies. How you plan to use the workstation will determine whether youll put
operating system,
memory, storage and graphics, he says. When these elements are your money into more cores, faster processors or more expensive GPUs. (See Making which can then
in balance, workstation users can explore more ideas in less the Case for Professional Workstations, Desktop Engineering, November 2014. schedule two
processes at once
time.
Put your dollars toward things that will directly impact performance, says David
RAID: An acronym
Waters, head of industry alliances, professional graphics, at AMD. Thats the biggest for redundant array
thing right from the beginning. of independent
disks, it is a method
of storing the same
data in different
Solid-State Drives places to improve
Every expert Desktop Engineering spoke to noted that if you are putting together a pro- input/output
performance and
fessional workstation, then youll have an SSD instead of a spinning hard disk prefer- fault tolerance.
ably for both the boot drive and storage. For the price (only a few hundred dollars more
than a spinning hard drive) you get a tremendous boost in performance.

The HP Z1 G2 is We try to go SSD when we can and when the budget allows, Chris Teague, a senior
an all-in-one
workstation aimed applications engineer at Solid Edge Reseller Saratech tells Desktop Engineering. You
at CAD users. Read can get a smaller SSD for the boot drive, and if you need a larger amount of space you
the full review can do a traditional hard drive for data. At least get an SSD for boot. If the budget
in the July 2014
edition of Desktop allows, then we have configured systems with four SSD drives in a RAID array for data,
Engineering. and a smaller one for boot.

Matthew Burke, principal and CEO, M2 Technologies, an Autodesk Reseller, is also


a fan of SSDs. In particular, FEA (finite element analysis) and CFD (computational
fluid dynamics) benefit a great deal with an SSD drive, he says in the August 2015 issue
of Desktop Engineering.1 SSDs have come down considerably in price so they almost
always tend to be worth the investment in a CAD workstation because they are simply
so much faster.

1
Answers to Your Workstation Questions, Desktop Engineering, August 2015

24 HPC Handbook | Chapter 2 | Building a Balanced Workstation Chapter 2 | HPC Handbook 25


According to Hamilton, users often under-buy when it comes to realistic rendering work. Those are the applications that will take advantage of more
AT A GLANCE: storage. CAD and simulation files are fairly large, and take a long cores because they are multi-threaded. Even for simulation, a single processor with six
SOLID-STATE DRIVES time to read and write, Hamilton says. What weve found in test- or eight cores will often suffice. For photo rendering, the number of cores improves
ing is that when you put in a faster storage system like an SSD, you performance exponentially. KEY WORDS
Price/performance ratio get much better CPU utilization. If you can get that data faster, you
now at a point that SSDs Turbo Boost:
get the processor going faster and see overall better performance. Simulation and rendering software applications are typically multi-threaded and can Intel Turbo Boost
are recommended for all therefore scale to multiple cores, Young says. Scaling ability can vary by applications, Technology
engineering workstations so users are advised to check with their CAD software vendor to verify the number of accelerates
processor
A smaller, less expensive FYI on CPUs cores recommended for any given workflow. and graphics
SSD can be paired with tra- For CPUs, speed is more important than core count if youre performance
for peak loads,
ditional spinning-disk hard running a CAD program. Look at single-thread performance Intels Parma says the Intel Xeon processor E5-1600 v3 Processor Family is designed for
automatically
drives to use as a boot drive instead of getting caught up in maximizing the number of cores. professional workstations that offer up to eight cores (and 16 threads using Intel Hy- allowing processor
if cost is a concern per-Threading Technology), four DDR4 memory channels, and up to 40 PCI Express cores to run
faster than the
The Intel Xeon processor E3-1200 v3 Processor Family is designed 3.0 lanes for excellent expandability and flexibility. For even more performance, work-
Can be configured in a rated operating
for entry workstations and brings workstation performance to stations equipped with the Intel Xeon processor E5-2600 v3 Processor Family offer up frequency if theyre
RAID array
designers, engineers and others who might not have historically to 36 cores (and 72 threads using Intel Hyper-Threading Technology), eight DDR4 operating below
power, current,
experienced this level of speed and capacity, says Parma. memory channels, and up to 80 PCI Express 3.0 lanes for maximum expandability and and temperature
It offers up to four cores (and eight threads using Intel Hyper- flexibility. specification limits.
Threading Technology), two DDR3 memory channels, and up to
16 PCI Express 3.0 lanes for expandability Workstations using the Intel Xeon processor E5-2600 v3 Processor Family are capable
and flexibility, he says. of mega-tasking for fast, efficient interaction with models while concurrently executing
simulations, renderings, ray tracing and modeling of the same system, Parma says.
Such an entry-level CPU may be all some
CAD users need. If you dont need a second processor or lots of cores, then moving down to a slower
processor with fewer cores can save a lot of money for CAD users. At a minimum, you
Dont overspend on the CPU, both in terms should have four cores (two for the OS, two for CAD). If you have other applications
of clock speed and the number of cores, running, another two or four cores might be in order.
Waters says. If you look at CAD or design
software, those CAD applications arent very Depending on the machine, you can also take advantage of Intels Hyper-Threading
multi-threaded at this point. In our per- technology or Turbo Boost, which provides dynamic control of the clock speed and core
formance analysis, CAD packages are still activation to improve performance in certain scenarios.
limited in terms of how many cores they can
get spinning when youre doing design work.
With the money you save on that, you can Buy a Lot of RAM
significantly upgrade the GPU. The good news when it comes to workstation configuration is that one of the most pow-
erful performance boosters is also one of the cheapest investments:
The number of cores is important as you add RAM. You need enough memory to avoid swapping hard drive space
other applications, and particularly if you to store and retrieve active data. Not having enough RAM will slow
plan to do any simulation, analysis or photo you down, even with a fast processor. You will always need enough AT A GLANCE: CPUs
system memory, especially as cheap as it is, it is always a good idea to Speed is more important
upgrade there, Waters says. than core count for CAD,
The DELL PRECISION TOWER 7910 can be ordered with but four cores is the rec-
up to a 10-core Intel Xeon processor and up to 64GB of However, the addition of RAM in 32-bit hardware will not neces- ommended minimum
RAM, along with multiple hard drive and graphics card sarily increase performance exponentially. Although the amount
choices. Core count is more im-
portant for multi-threaded
simulation, analysis and
rendering applications

26 HPC Handbook | Chapter 2 Chapter 2 | HPC Handbook 27


RAM CONFIGURATIONS If you get into assemblies that are very, very large, upwards of 20,000 components or
more, then you may need 24G or 32G of RAM, says Adrian Fanjoy, technical services
manager at CATI. For pretty much everything else (in SOLIDWORKS), once you get
SECONDS TO COMPLETTION to 16GB youre not going to get much more improvement.

The GPU
The size and complexity of the models are the biggest
consideration when it comes to GPU selection. The system
should provide a 15 frames per second (fps) to 20 fps re-
fresh rate; anything below that would look slow or clunky,
although what level of performance is good enough is a
subjective measure.

For entry-level workstations, many applications are certified


to use Intel HD Graphics P4600 graphics, says Intels Parma.
For many entry workstation users, using Intel HD Graphics
P4600 graphics with an Intel solid-state drive (SSD) will
AVAILABLE RAM result in a more balanced workstation and a better overall
user experience than using an add-in graphics card with a
RAM & PERFORMANCE CATIs research shows anything beyond 16GB of RAM is not necessary to improve traditional hard disk drive (HDD), he says.
SOLIDWORKS performance. However, RAM is inexpensive and moving to 32GB of RAM could help other
computing processess. Learn more here. However, a discrete graphics card is recommended by many
workstation software vendors, and is especially important
when using an expert workstation or professional worksta- ECC, FYI
tion, as the processors in these workstations do not include Error-correcting code (ECC)
of RAM is crucial for the overall machine performance, it by itself cannot guarantee integrated graphics. For complex geometries, higher-end RAM protects workstations from
maximum performance, says Muthukumaran Kanniappan, head of delivery at Bar- graphics cards can significantly boost the refresh rate. The potential crashes and changes in
ry-Wehmiller International. Choosing the right RAM capacity to suit other compo- money you spend on the GPU will be evident in the interac- data by detecting and correcting
nents of the hardware can help enhance performance. Users need tivity of the design software as the models and assemblies get single-bit memory errors.
not always aim for the maximum, because you can upgrade the bigger, and you get involved in more simulation and render- According to research by Google
RAM later. ing. For users not planning to design or visualize 3D graph- and the University of Torontos
AT A GLANCE: RAM ics, then the integrated graphics available on the processor Department of Computer Science,
RAM is cheap Put as much RAM in as your budget will allow, reinforces Burke2. might be sufficient. one in three workstations have
Certainly there are published guidelines available for both min- at least one correctable memory
It allows you to avoid
imums and recommendations, but they generally dont take into Another GPU consideration: The use of GPU-accelerated error per year. In workstations
swapping, which slows
account the other applications that are going to be up and running. applications that can access the compute cores in the card without ECC RAM, single-bit
performance by using hard
for computational tasks. If youre using that type of software, memory errors can not only crash
drive space to store and
SOLIDWORKS reseller Computer Aided Technology (CATI) has youll need a higher-end graphics card to support it. a workstation and corrupt data,
retrieve active data
done benchmarking on professional workstations for several years, but can sometimes change data
RAM does not increase and in the companys recent tests3 found that 16G of RAM was suffi- If you cant spend a lot of money on the GPU, you can also without being detected.
performance on a one-to- cient for most models, regardless of size. consider making adjustments to the design software. You can
one basis lower the level of detail in a model and get smooth rotation, it ECC RAM is compatible with Intel
16GB is a good rule of 2
Answers to Your Workstation Questions, Desktop Engineering, August, 2015 just depends on the comfort level the engineer has with that Xeon processors.
thumb for CAD, and can be 3
Maximize SOLIDWORKS Performance, Desktop Engineering white paper spon- trade-off and how much detail is required. There are also mid-
increased as needed sored by Intel, July 2015

28 HPC Handbook | Chapter 2 Chapter 2 | HPC Handbook 29


range cards available that can provide a lot of horsepower, even for
larger models.
IF NOT NOW, WHEN?
AT A GLANCE: GPUs
For CAD users, the graphics processor is far less important than With all of the evidence in favor of work- prevent data corruption; and the doubling of
Consider GPUs if you are having fast application performance, Kanniappan says. After all, if one station upgrades, why are many design disk drive speeds thanks to SSDs. Virtualization
working with large, com- spends 10% of the day rendering and 90% designing, the computer engineers still limping along on three-year-old where one workstation can run multiple
plex models should be set up to optimize productivity on CAD applications. But (or older) computers? operating systems or be divided to handle
Integrated graphics are having said that, if the application of a particular machine is specifically different tasks is becoming increasingly
often sufficient if not work- for inspection, collision checking or drawing quality control, then it Some of it is budgetary. Those who hold the common, as is the use of multiple monitors.
ing with 3D graphics makes perfect sense to invest in a quality graphics processor. In the end, purse strings may not have bothered to calcu-
it all boils down to the end-application of the machine. late the quick ROI of an upgrade, or may not fully Processors, bus speeds, graphics cards and hard
understand that engineers have different com- drives are all faster, says Lenovos Tom Salo-
puting needs than their co-workers in account- mone. Engineers may think that hardware is just
Bigger, Better Displays ing or sales departments. a little faster or the same, but the truth is it all
Another element that can drive GPU selection is the display. Most CAD users have adds up to let them work much, much faster.
more than one monitor, a practice that improves productivity. If you have multiple Other reasons for keeping engineers from the
screens, then youll need professional graphics card like the AMD FirePro or computing power they need are cultural. People Source: Workstations: Now is the Time to Upgrade,
Desktop Engineering white paper sponsored by Lenovo.
NVIDIA Quadro line that can easily drive multiple monitors. The Intel Xeon processor are comfortable with the status quo because
E3-1200 v3 Product Family supports up to three displays (see pages 11 and 31-36 they dont know what they are missing.
of this document). Each display port is capable of driving resolutions up to 3840x2160 Some are reluctant to make a change
at 60 Hz through DisplayPort and 4096x2304 at 24 Hz/2560x1600 at 60 Hz using during a large project and theres
HDMI. always a large project.

KEY WORDS There are also monitors available with 4k and 5k resolution and wide aspect ratios. As To get beyond the budgetary and cultur-
4k and 5k: you increase the resolution of the display, you are putting the same sort of pressure on al shortsightedness, look to the future.
resolutions of the GPU to draw that many pixels. Models are going to continue to grow in
40962304 and
5120x2880 (16:9
size, and modern design software capa-
aspect ratio), We now have GPUs that have so much rasterization that we can drive two or three bilities will continue to take advantage of
respectively. 4k displays off a single GPU, Waters says. But you cant get that performance from the latest advances in computing hard-
entry-level products. ware.
Rasterization:
converting an
image into pixels/ One display issue that sometimes gets overlooked is the signal range of the display. The Over the last three years, those hardware
dots for display on a bandwidth connection on that cable has a lot of data going across it, Waters says. You advances include more productive
video display
have to be careful with lower-cost displays, because they may only be able to send a 2k computational and graphics processing
Upconvert: scaling signal and then upconvert it. You may struggle with some instability. via new architectures and more cores;
a lower resolution faster memory speeds with error-
signal to fill a higher correcting code memory that helps
resolution screen
Software Optimization
The last piece of the workstation optimization puzzle isnt part of the hardware at all.
Software settings can provide another big boost to performance, although exactly what
those settings are will vary by application.
The LENOVO THINKSTATION P300
SFF WORKSTATION is targeted at
Some OEMs can help automate the process of matching the software to the machine. entry-level 3D and 2D CAD users.
Dell, for example, has a performance optimizer on its Precision workstations that Read the full review.

30 HPC Handbook | Chapter 2 Chapter 2 | HPC Handbook 31


automatically tunes the hardware and settings to match specific workstations that can be shared across loca-
AT A GLANCE: types of applications. HP, similarly, offers Performance Advisor tions.
DISPLAYS for optimization settings on its HP Z Workstations and mobile
workstations. Other applications may take advantage of features found
Multiple monitors
on specific CPUs or GPUs. AMD, for instance, has co-engineered Return on Investment
improve productivity
different features with software vendors that can also provide How often you update your hardware will also
4k and 5k displays allow performance benefits in certain configurations. play a role in workstation component selec-
you to see more details, tion. If you wont be investing in a new ma-
but require more graphics Workstation configuration is always limited by the budget, and that chine for three to five years, you might want to
processing muscle means companies will need to make trade offs. Those trade offs should buy more horsepower than you need now in
Avoid displays that will be be based on application usage. anticipation of needing it a few years down the
forced to upconvert signals road as you cycle through software upgrades.
For example, investment in high-end plotting and printing
solutions may depend purely on the need for inspection, record keeping, approvals and The general rule of thumb (for investing
presentation, Kanniappan says. And such investments will also vary greatly depending in a new workstation) is three years, but that
on the organizations expectation of the deliverables of the engineering function. can depend on a lot of things, says Teague.4
Did you buy at the high end three years ago,
The availability of cloud or virtualization options could mitigate the need to invest in or did you get a cheap workstation? After
on-premise rendering of simulation capabilities, or allow you to consolidate higher-end three years look around and see what the
technology is now vs. what you bought. Even
if the hardware isnt obsolete, a fresh install of
Windows might make things faster.

ROI CALCULATION EXAMPLE Selecting the most practical workstation with the BOXX TECHNOLOGIES overclocks the processors in its workstations, which
right features to match your application usage can boost productivity in applications like CAD. Read the full review.
TIME SPENT DOING ANNUAL BENEFIT FROM can provide hours of increased productivity. The effi-
DESIGN WORK: INCREASING DESIGN ciency gains should be top of mind, particularly for organizations that are considering cutting
A third of 250 working days PRODUCTIVITY: corners on the hardware in order to reduce costs.
= about 83 days per year $35 per hour for 8 hours
a day for 29.88 days According to Desktop Engineerings series of reports benchmarking various simulation FOCUS ON PRODUCTIVITY:
ANNUAL COST OF = $8,366.40. THE INVESTMENT IN A WORK-
and design software by comparing three-year-old software running on three-year-old
STATION IS INSIGNIFICANT
AN ENGINEERS hardware against their current equivalents, engineers are seeing simulation time savings COMPARED TO THE COST
DESIGN WORK: ADDITIONAL TIME of up to 6X by upgrading.5 OF THE ENGINEER
$35 per hour for 8 hours SPENT DOING DESIGN WORK USING IT.

a day for 83 days BY BEING 36% MORE


Companies are trying to be cost conscious, but you have to look at the big picture,
= $23,240 PRODUCTIVE:
Hamilton says. The investment in the workstation is pretty insignificant compared to
83 days x 36%
= 29.88 days the cost of the operator using it. If you do the ROI analysis with that in mind, making
ROI OF AN M2800 MOBILE your operator more productive means the workstation will pay for itself in about six
WORKSTATION months. You have to think about making the engineers as productive as possible, not just
THAT COSTS $1,079:
saving a few bucks.
About six weeks
4
Answers to Your Workstation Questions, Desktop Engineering, August 2015
Source: Making the Case for Professional Engineering Workstations, 5
Simulation for the Masses, a Desktop Engineering benchmarking report, sponsored by Intel,
Desktop Engineering, November 2014. Dell and COMSOL. Focus on Faster Mecahnical Simulation, a Desktop Engineering benchmarking reports,
sponsored by Intel, Dell and ANSYS.

32 HPC Handbook | Chapter 2 Chapter 2 | HPC Handbook 33


CHAPTER 3:
PARALLELIZATION
PRIMER
Parallel processing has revolutionized
the way design software operates.

42 HPC Handbook | Chapter 3 Chapter 3 | HPC Handbook 43


F
ew technology advances have had quite the same impact on design
software as the advent of parallelization or parallel processing. By taking
COPROCESSORS AND PARALLELISM
advantage of the power of multiple cores and multiple CPUs (central The Intel Xeon Phi is a coprocessor, not a general purpose applications (not just graphics
processing units) and GPUs (graphics processing units), engineers and central processor. Just as AMD and NVID- applications) that take advantage of the GPUs
designers have been able to quickly and cost-effectively wrangle very IAs GPUs are graphics coprocessors, Phi, too, is parallel processing power. This allows comput-
large simulation and rendering tasks that previously would have required a coprocessor. In the case of Phi, however, its ing intensive applications, like simulation, to run
outsourcing and many hours (or days) of delays. not restricted to graphics tasks. Its a general on GPU clusters.
purpose coprocessor, aimed at handling highly
Both AMD and NVIDIA must have realized asking
Parallelism leverages concurrency to gain better performance. Parallel computing parallel computing jobs (simulation and render-
programmers to master a whole new program-
defines these very large simulation problems in smaller pieces and simultaneously solves ing are the top two types for engineers). Though
ming language to write parallel-processing code
each piece using multiple processors or multiple computers. As a result, you can get originally developed as graphics coprocessors,
for their products would put a roadblock on the
KEY WORDS
your answer faster (lower latency) or you can find more answers in the same amount of AMD and NVIDIAs GPUs are also well on the way
race to HPC. So both are working hard to make
Parallelism: time by increasing throughput. to tackle general purpose computing jobs.
their respective architecture HSA and CUDA
Cooperatively
solving large In the case of AMD, the path to parallelism is its easier to work with for programmers who use
computations If you are handling and processing a lot of data, you are ripe for using parallelism, says heterogeneous system architecture (HSA). Its a common programming languages, like C, C++,
by dividing them James Reinders, parallel programming evangelist at Intel. mouthful, but basically, its a computing envi- JAVA, Fortran, and Python.
among multiple
processors, which ronment that fuses CPUs and GPUs functions
Parallelism provides an opportunity to direct the multiple functions of a computer In that respect, Intels Phi may have an advan-
solve the smaller into a single device (the root of AMD Fusion, as
computations system in a fashion similar to conducting an orchestra. How can the system use multi- tage over its rivals. Intels Many Integrated Core
the product line is called). The HSA Foundation, a
simultaneously.
ple resources in concert, rather than one a time? You have to think about how you are (MIC) architecture is basically a small cluster of
nonprofit industry group, describes HSA as par-
using the total computer system, Reinders says. CPU cores in a single chip. Therefore, program-
Hyper-Threading: allel computation utilizing CPU, GPU and other
mers are expected to be able to write parallel
Intel Hyper- programmable and fixed function devices.
Threading This is different than multi-threading, in which a single CPU or core executes multi- code for Phi using standard programming lan-
Technology ple processes or threads concurrently by splitting up the data and tasks into sub-tasks Taking a slightly different path, NVIDIA proposes guages compatible with CPUs.
enables multiple CUDA as the programming environment to write
independent on shared memory. On a CPU with Hyper-Threading, for example, you can perform Kenneth Wong
instructions to multiple sequential tasks by running several dozen threads in parallel with multiple
run on each core.
cores. Multi-threading is for in-process computation, while parallel computing is for
One physical COMPLIERS
core appears as out-of-process computations, says Silvina Grad-Freilich, MathWorks senior manager SOURCE LIBRARIES,
two processors for parallel computing and deployment products. PARALLEL MODELS
to the operating
Intel MIC
system, which can architecture
then schedule Parallel processing can be enabled via modern CPUs or GPUs. The CPU is good at coprocessor
two processes at loading and holding data so that the same data can be used over and over again. GPUs Multicore CPU Multicore CPU
once, speeding the
completion of the
can perform parallel computations that are the same or similar on different data.
instructions.
GPU accelerators complement CPUs to provide the best app performance for end
users, says Will Ramey, senior product manager of accelerated computing at NVIDIA.

Faster, Better Results


With software applications that have been designed to take advantage of parallelism,
engineers can complete complex tasks much faster.
PROGRAMMING MODELS Intels vision is to span from multicore to many-core architecture with consistent programming
With parallelized software, our engineering customers can really achieve enhanced models, languages, tools and techniques. Vendors can write one program for both a multicore CPU and the Many Integrated
Core (MIC) architecture in other words, first- and second-generation Intel Xeon Phi products. Image courtesy of Intel.
engineering productivity by accelerating simulation throughput, says Wim Slagter,

44 HPC Handbook | Chapter 3 | Parallelization Primer Chapter 3 | HPC Handbook 45


lead product manager for high performance computing at ANSYS. It also helps them
to make more efficient product development decisions.
THE BLURRED BOUNDARY BETWEEN CPU AND GPU
Engineers can also achieve higher fidelity insight into product performance that With the launch of new processors with than graphics. To do so, their software has to sup-
couldnt be gained any other way. Parallelized software capabilities allow engineers to integrated Iris Pro graphics, Intel begins port NVIDIAs CUDA parallel processing language.
simulate larger models and more complex models, so that more accurate design deci- describing its products as more than CPUs. The
KEY WORDS Intels Blakely suggests software vendors can
sions can be made throughout the design or development cycle, Slagter adds. Xeon E3-1200 V4 is a CPU and GPU on the same
tap into the additional horsepower in Iris Pro-
Amdahls Law: chip, said Jim Blakley, general manager of Visual
Named for based CPUs with little or no additional work. He
If the simulation runs faster, then you can do more iterations of exploring parameters Cloud Computing for Intels Data Center Group.
computer architect explained, If the application is written to use
Gene Andahl, in of the design space much faster, Ramey says. You can refine those designs faster, Iris Pro is a major part of Intels strategy to drive OpenGL or DirectX, then Iris Pro work without
parallel computing
it is used to describe
which results in better products. growth in traditional GPU market segments. To any additional programming effort. NVIDIA
the theoretical speed up graphics-heavy workloads (like ray- also supports OpenGL and DirectX. These are
speed-up of traced rendering) and compute-intense workloads graphics programming APIs. If you want to do
applications using (like simulation), engineering and design software things that are more compute-oriented, then we
parallelism. Because
Where Core Count Matters
some processes Where this approach to computation matters most in design is typically in simulation users tend to rely on the GPU as a coprocessor to support OpenCL programming on the GPU. This
are inherently and rendering involving large amounts of data. Computational fluid dynamics (CFD) augment the CPU. The GPUs highly parallel nature is the direct comparison with CUDA. Whereas
serial, there is a makes it an attractive device for such applications. CUDA is proprietary NVIDIA programming mod-
point at which the simulations, for example, benefit from higher core counts. For a large aerospace compa-
However, Intels CPU with integrated graphics can el, OpenCL is industry standard.
application cant ny, huge simulations rely on throwing large numbers of cores at the computations.
disrupt the market dynamics, especially for the
be made any more The applications that benefit the most are those that have computations heavy enough For highly parallel applications, computer users
parallel. That point low-end and mid-range segments.
varies depending that they can benefit from breaking the data into chunks, says Grad-Freilich. You are expected to continue to rely on NVIDIA Tesla
on the type of work need to make the right decision for the technology that you use based on your actual Blakeley clarified, Iris Pro graphics is not meant GPU or its equivalent. However, if software pro-
involved. application needs. to compete in the high-end GPU market. The grams can take advantage of Iris Pros additional
performance of Iris Pro is roughly equivalent of a horsepower without much effort, its set to chal-
OpenGL: Open
mainstream graphics card. GPU maker NVID- lenge the GPUs dominance in visualization-re-
Grpahics Library However, not every task within an application can be parallelized. Companies may not
is an environment IA, for instance, offers its NVIDIA Tesla cards lated tasks like image editing and 3D modeling.
see the speed-up they expect if there are lengthy sequential processes involved in the
for developing to the high-end market, and Quadro cards to
portable, interactive solution. For example, an application may have a sequential part that runs before and Workstation users may also split their workload
mainstream market, which encompasses CAD
2D and 3D graphics after the other computational work, so the overall speedup of the application is less than on the two GPUs in the same machine. Blakely
applications.
software users.
linear because of the serial nature of the beginning and end processes, Grad-Freilich pointed out, You can use one of those GPUs for
DirectX: A set says. In the last four to five years, some researchers and graphics, another for compute-intensive work.
of low-level scientists began using the GPU for much more
application
Kenneth Wong
programming That limit on acceleration is explained by Amdahls Law, which limits the theoretical
GRAPHICS PERFORMANCE VIDEO STREAM PERFORMANCE
interfaces (APIs) speed-up of applications using parallelism. Because some processes are inherently serial,
that provides
there is a point at which the application cant be made any more parallel. That point 1.4X4
Windows programs MORE
varies depending on the type of work involved. 1.8X PERFORMANCE2
with hardware-
accelerated OVER STREAMS
multimedia support. 4,300 PER RACK2
There are software applications that dont benefit from parallelism, or that dont ben-
APIs: Application efit from an exponential increase in core count. Mechanical simulations, for example,
programming do not scale up to tens of thousands of cores, Slagter says. Weve shown in our latest
interfaces make
it possible for release that it scales up to 128 cores for a whole suite of different benchmarks.
applications to
E3-1200 v3 w/HD P4600 E3-1200 Professional Graphics Performance
interact and share If you have data that must be operated on sequentially, and theres only a small amount
data. E3-1200 v4 w/Iris Pro Graphics P6300 E3-1200 v4 w/Iris Pro Graphics P6300
of computation for each piece of data, that application would not benefit from paral-
OpenCL: A lelization, Ramey adds. The good news is that when smart programmers get engaged,
standard for cross- GRAPHICS PERFORMANCE The E3-1200 v4 with Iris Pro Graphics provides an overall boost to graphics and video
platform, parallel
the algorithms in applications that are fundamentally serial can often be redesigned so stream performance, which can help transform engineering workstation and multimedia workflows.
programming of they can run in parallel.
processors.

46 HPC Handbook | Chapter 3 Chapter 3 | HPC Handbook 47


Modern design workstations typically have multiple processors, which provides HPC- Limits of Parallelism
like capabilities on the desktop. Software of all types has been designed to take better With the availability of cloud computing resources, engineers can access tens of thou-
advantage of those capabilities. sands of cores (or more) for very large simulations.

We keep making our data problems bigger and bigger, Reinders says. Because of But how far can this go? Is there an upward limit at which simply adding more cores
that, computers can get more use from parallelism as time goes on. As long as we use isnt going to provide any more value?
more and more data, then parallelism will be required.
In theory there is no upward limit. There are super computers, for instance, that use
SOFTWARE MUST BE
DESIGNED TO TAKE The majority of engineering and design software has been architected to take advantage tens of thousands of nodes. However, there is a point at which particular problems or
ADVANTAGE OF PARALLEL of parallelism. Some software providers have struggled with building tools to take ad- applications no longer see a return on the investment in computational horsepower.
PROCESSING.
vantage of this approach because of inexperience or because the software was originally When you look at simulations, there is a level at which you can break the problem
designed when single-core was the standard. down to, where you are still getting good bang for your buck, Reinders says. If you
keep breaking it down into smaller pieces, there can be too much overhead involved in
The question becomes how much does this perturb the way the program was written orig- moving the data around.
inally, Reinders says. There are a lot of applications that were written when there was just
one core in a machine. This can affect the architecture of the application in a profound way. According to Amdahls Law, once you have infinitely fast computation, the bottleneck
is going to move somewhere else, Ramey adds. Total execution time might be limited
Developing software that takes advantage of these capabilities also requires sustained and by the storage system. Another area might be communication between nodes, when
continued HPC software development to effectively leverage the hardware. The soft- adjacent nodes need to exchange some state information. At the extreme end, with hun-
ware should be able to support simulation resources where they are located, Slagter says. dreds of thousands of cores, its possible that the bottleneck is going to be the speed of
Certifying remote software solutions is important, along with expertise and support from the interconnect or bandwidth between nodes.
HPC partners like Intel, NVIDIA, Hewlett Packard, etc., so you can make sure that you
have an optimized reference architecture for the software and good support. Still, with access to thousands of cores, there is a definite payoff in performing more
simulations at the same time. But if there is only a limited amount of data you want to
process, then there will be a limit to how many cores you can throw at the problem and
see a noticeable difference in how fast the problem is solved, Reinders says.

With cloud-based resources, uploading and downloading data and results can create
bottlenecks as well. Its very clear that the cloud, much like on-premise HPC infra-
CALCULATION 1 structure, is going to provide high bandwidth, storage, fast processors and low latency,
Slagter says. But the cloud may put a limit on the data because the huge amount of data
can affect the performance of the system. The time required to move files and download
CORE 1 results can take many hours. That may cause a productivity hit and impede the use of
CORE 2 CALCULATION 2 HPC in the cloud.
DATA DATA
STREAM STREAM
CORE 3 CALCULATION 3 The solution is to leave the data where it is and perform operations in the cloud. There
are specific requirements for cloud computing in terms of job scheduling to optimize
CORE 4
use of cloud infrastructure, Slagter says. Its important that you take care of data man-
agement and simulation job management.
CALCULATION 4
Parallelism has helped engineers improve and accelerate large computational tasks, with
very little downside. As software vendors continue to take advantage of these advances
in CPU and GPU capacity, simulation and rendering work will continue to be faster
and more responsive, helping designers to increase iterations and improve design.
PARALLEL PROCESSING In parallel computing, multiple cores can be used to cooperatively solve large
computations by dividing them into smaller computations that can be solved simulataneously, thus
speeding the overall solution time.

48 HPC Handbook | Chapter 3 Chapter 3 | HPC Handbook 49

Você também pode gostar