NERSC Managers Shed Light on ‘Edison’

By Nicole Hemsoth

April 29, 2013

At the end of March, the Department of Energy’s National Energy Research Scientific Computing (NERSC) Center accepted the first phase of its new Cray Cascade system. Named Edison, the new machine is a Cray XC30 with 664 compute nodes and 10,624 cores. Each node has two eight-core Intel “Sandy Bridge” processors running at 2.6 GHz (16 cores per node), and has 64 GB of memory.

While the user environment on Edison is remarkably similar to that on Hopper, NERSC’s current main system, a number of new features and technologies are available on the Edison Phase I system, including the Cray Aries high speed interconnect, Hyper-Threading technology, the Sonexion storage system, and an external batch server. Once fully installed later in 2013, Edison will have a peak performance of more than 2 petaflop/s. The integrated storage system will have more than 6 petabytes (PB) of storage with an I/O bandwidth of 140 gigabytes (GB) per second.

Bucking the trend toward GPUs and hybrid architectures, Edison will use Intel processors exclusively. To find out the reasoning behind the design and deployment of Edison and what it means to NERSC’s 4,500 users, Jon Bashor of Berkeley Lab Computing Sciences spoke with NERSC Division Director Sudip Dosanjh, NERSC Systems Department Head Jeff Broughton and Advanced Technologies Group Leader Nick Wright.


Jon Bashor: First of all, how did the system come to be named Edison?

Jeff Broughton: At NERSC, we name our computers after famous scientists, like Hopper for Grace Hopper and Franklin for Benjamin Franklin. In this case, we were looking for someone iconic, someone who represented American team science. Thomas Edison was the obvious choice, especially for his work in the field of energy. His work had practical applications, and while NERSC supports mainly basic research, naming the system after Edison will be a constant reminder to consider the applied connotations of the science we support.

Bashor: When the procurement was announced in June 2012, the new system was described as delivering more than 2 petaflop/s peak performance, about twice that of Hopper. When other centers announce procurements, the new systems often provide an order of magnitude increase in performance. Is NERSC exiting the peak performance race?

JB: Although NERSC systems in the past have been highly ranked, NERSC has never been part of the peak performance “race.” We have always focused on the sustained performance of scientific applications as the basis for our procurements. For the most part, over the past 10 years, sustained performance has been a relatively constant fraction of peak performance. But with the move to GPUs and manycore systems, it has become increasingly difficult to maintain the traditional sustained/peak ratio. Getting users to move to these new architectures requires a major effort to port their applications.

As we transition to the path to exascale over the next decade, we think we can ask our 4,500 users to make the move to an entirely new class of architecture only once. We made a conscious decision that this is not the time for our users to make the transition and to let the architectures get sorted out before making that move. Our next system, now referred to as NERSC-8, will adopt one of the new energy-efficient architectures.

We are also seeing in our user community a big increase in data-intensive computing, with a focus on high throughput and single-node performance, so another goal was to provide a system that could meet the needs of data-intensive applications while continuing to support conventional HPC.

And we will continue to apply one of our unofficial benchmarks – the number of published research articles published by our users based on calculations performed at NERSC. We’ve averaged 1,500 per year for the past five years – another example of the sustained performance we are most interested in.

Bashor: Edison will be one of the first Cascade systems delivered by Cray. Is there any risk associated with this? Why did NERSC take this route?

Nick Wright: Over the years NERSC has taken calculated risks on numerous occasions by deploying a low serial number version of a system. This goes as far back as 1978, when a Cray-1 with serial number 6 was deployed, and 1985, when NERSC installed the first Cray-2 machine. Other examples include the T3E-900 installed in 1997, which was the largest I/O system built to date, and Hopper, one of the first Cray XE6 machines, deployed in 2009. There are significant benefits for the NERSC user community in terms of increased scientific output from having access to the latest, highest-performing technology. Also, there are significant benefits to NERSC and its sponsors because utilizing the latest technology maximizes the useful lifetime of a system and the return on investment. In this case, the Cray Cascade system has both the latest Intel processors and the highly scalable Aries interconnect, both of which are already delivering exceptional performance to our users.

JB: Additionally, Edison has a novel cooling scheme that will allow us to move it to the Computational Research and Theory (CRT) Center, our new facility now under construction. CRT will use “free” cooling using only outside air and cooling towers. Thanks to our Bay Area climate, we can run this way year round without the added cost of mechanical chillers. Cascade is designed to operate with the warmer air and water temperatures that we can expect during certain parts of the year.

Edison has proven to be very reliable right out of the starting gate. Phase 1 of the system was installed in late 2012, and has already proven very popular with our users.

Bashor: Not only is this the first Cray machine to use Intel processors, but it also incorporates the new Aries interconnect and the Cray Sonexion storage system. What will this mean for NERSC’s 4,500 users?

Sudip Dosanjh: We believe Edison will be a very productive system for our users. It has very high memory bandwidth, memory capacity and interconnect bandwidth relative to its compute capability. Data movement is the limiting factor for a large fraction of the 600 codes that run at NERSC. Floating point units are often idle waiting for data to arrive. Many codes spend a few percent of their total runtime performing floating point operations, the rest of the time is spent accessing memory or calculating memory addresses. Edison will be a very effective platform for running the very broad range of science codes at NERSC.

JB: In short, Edison will be a very scalable and robust system for the broad range of users that NERSC has. We support more than 700 projects across all the program offices of the Department of Energy’s Office of Science: Advanced Scientific Computing Research, Basic Energy Sciences, Biological and Environmental Research, Fusion Energy Sciences, High Energy Physics and Nuclear Physics. The Intel processors and Aries interconnect have delivered good performance on a diverse class of algorithms.

Since the programming and operating environment is essentially the same as Hopper, it’s been extremely easy for users to move codes from one system to another, even on a day-to-day basis. It just takes a simple recompile to optimize performance.

NW: NERSC users are frequently performing science runs at scale, both to generate science results as well as explore the capabilities and limitations of their codes. The capabilities of the new highly scalable Aries interconnect will enable greater performance at scale and will deliver significant sustained performance.

Bashor: The system also provides 64 GB of memory per node. How is this different from other systems and what’s the benefit to users?

NW: We could have purchased Edison configured with 32 GB per node and spent the money saved on boosting the peak flops of the system. As Jeff said though, our focus is on enabling scientific productivity and we felt the benefits to users from extra memory far out weighed those that a slightly higher peak flop rating would have delivered. The 64 GB per node is twice the amount of memory we have per node in Hopper. This will allow those codes that are more data-intensive to process more quickly and to solve larger problems.

Bashor: NERSC’s future home, the CRT center currently under construction in Berkeley, will use the Bay Area’s “natural air conditioning” to cool the machine room and improve efficiency. What’s being done in the meantime?

JB: At our current facility in Oakland, we have essentially the same climate, and we are prototyping the CRT cooling infrastructure using Edison. We removed 1,100 tons of chillers and replaced them with heat exchangers. Cooling comes from evaporation in cooling towers exclusively. This gives us an opportunity to test the principles we are implementing and adjust the CRT design as needed. The expected PUE (power usage effectiveness) for Edison is approximately 1.1, which represents a two-thirds savings in the energy costs for cooling compared to similar systems cooled with mechanical chillers.

Next >> Looking Farther Ahead

Bashor: Looking farther ahead to exascale, NERSC’s 4,500 users and their 600 applications will face some real challenges to maintain their current scientific productivity of 1,500 papers each year. How will you manage the transition to exascale systems from the users’ perspective?

SD: Exascale computing will impact all scales of computing from supercomputers to racks because the fundamental building blocks, processors and memory, are changing dramatically. Clock speeds are expected to remain near a GigaHertz for the foreseeable future, and the concurrency in processors is expected to increase at a Moore’s Law pace. All of the codes running at NERSC will need to transition to energy efficient architectures during the next few years. If we can’t make this leap, users will be stuck at today’s performance levels and we will miss many opportunities for scientific discovery. We will work with our users to make this transition as smooth as possible. We have started an application readiness effort to begin transitioning our codes in the NERSC-8 time frame.

JB: As I mentioned, our plan is to start transitioning users to exascale programming models – which will involve increased parallelism and non-transparent memory hierarchies – with our procurement of the NERSC-8 system, which will arrive in late 2015. We expect this system to be some variety of a GPU or manycore architecture.

Of the 600 applications running on NERSC systems, we know that just 10 of them account for about 50 percent of the time used on the machines. We will make a concerted effort to port those codes to NERSC-8. There will still be a number of applications that will be difficult or not cost-effective to immediately transition to the new machine, so they will stay on Edison and be given more time to transition. We see Edison as a kind of safety net for users of those applications.

Bashor: Last question. NERSC recently completed its 10-year strategic plan. Can you give us a short summary of where the center expects to be in 10 years?

SD: The aggregate computing needs of Office of Science research teams at NERSC will be well into the exascale regime by the end of the decade. Science teams need to run simulations at hundreds of petaflop/s, and they need to run thousands to millions of petascale simulations. We will deploy pre-exascale systems in 2016 and 2019. We anticipate deploying our first exascale system, NERSC-10, in 2022.

We are also seeing the growing importance of big data and many recent scientific breakthroughs enabled by NERSC involve large data sets. For the past four years, our users have imported more data to NERSC to analyze than they have exported. Many months, we import more than a petabyte of data! We will begin enhancing our data capabilities starting in 2014, and we will deploy data systems, storage, advanced networking, and enhanced user services so that current users and DOE experimental facilities can move and process exabytes of data early in the next decade.

While we have talked a lot about systems in this interview, I want to close by talking about the most critical component of NERSC’s very successful 39-year history – our staff. We run a very effective operation, operating Edison, Hopper, three large clusters, a 20-petabyte HPSS data archive and more, all stitched together seamlessly by the NERSC Global Filesystem. On the services side, we providing extensive technical support to our 4,500 users as well as providing the intellectual infrastructure needed to advance scientific discovery across DOE’s research mission areas. Our users are also very complimentary of our staff – consistently giving us excellent scores in our annual survey. Achieving this requires both extensive expertise and true dedication on the part of our staff. Although I only joined NERSC last fall, I am very proud to be part of such an accomplished organization.

About NERSC and Berkeley Lab

The National Energy Research Scientific Computing Center (NERSC) is the primary high-performance computing facility for scientific research sponsored by the U.S. Department of Energy’s Office of Science. Located at Lawrence Berkeley National Laboratory, the NERSC Center serves more than 4,000 scientists at national laboratories and universities researching a wide range of problems in combustion, climate modeling, fusion energy, materials science, physics, chemistry, computational biology, and other disciplines. Berkeley Lab is a U.S. Department of Energy national laboratory located in Berkeley, California. It conducts unclassified scientific research and is managed by the University of California for the U.S. DOE Office of Science. For more information about computing sciences at Berkeley Lab, please visit www.lbl.gov/cs.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industry updates delivered to you every week!

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing power it brings to artificial intelligence.  Nvidia's DGX Read more…

Call for Participation in Workshop on Potential NSF CISE Quantum Initiative

March 26, 2024

Editor’s Note: Next month there will be a workshop to discuss what a quantum initiative led by NSF’s Computer, Information Science and Engineering (CISE) directorate could entail. The details are posted below in a Ca Read more…

Waseda U. Researchers Reports New Quantum Algorithm for Speeding Optimization

March 25, 2024

Optimization problems cover a wide range of applications and are often cited as good candidates for quantum computing. However, the execution time for constrained combinatorial optimization applications on quantum device Read more…

NVLink: Faster Interconnects and Switches to Help Relieve Data Bottlenecks

March 25, 2024

Nvidia’s new Blackwell architecture may have stolen the show this week at the GPU Technology Conference in San Jose, California. But an emerging bottleneck at the network layer threatens to make bigger and brawnier pro Read more…

Who is David Blackwell?

March 22, 2024

During GTC24, co-founder and president of NVIDIA Jensen Huang unveiled the Blackwell GPU. This GPU itself is heavily optimized for AI work, boasting 192GB of HBM3E memory as well as the the ability to train 1 trillion pa Read more…

Nvidia Appoints Andy Grant as EMEA Director of Supercomputing, Higher Education, and AI

March 22, 2024

Nvidia recently appointed Andy Grant as Director, Supercomputing, Higher Education, and AI for Europe, the Middle East, and Africa (EMEA). With over 25 years of high-performance computing (HPC) experience, Grant brings a Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing po Read more…

NVLink: Faster Interconnects and Switches to Help Relieve Data Bottlenecks

March 25, 2024

Nvidia’s new Blackwell architecture may have stolen the show this week at the GPU Technology Conference in San Jose, California. But an emerging bottleneck at Read more…

Who is David Blackwell?

March 22, 2024

During GTC24, co-founder and president of NVIDIA Jensen Huang unveiled the Blackwell GPU. This GPU itself is heavily optimized for AI work, boasting 192GB of HB Read more…

Nvidia Looks to Accelerate GenAI Adoption with NIM

March 19, 2024

Today at the GPU Technology Conference, Nvidia launched a new offering aimed at helping customers quickly deploy their generative AI applications in a secure, s Read more…

The Generative AI Future Is Now, Nvidia’s Huang Says

March 19, 2024

We are in the early days of a transformative shift in how business gets done thanks to the advent of generative AI, according to Nvidia CEO and cofounder Jensen Read more…

Nvidia’s New Blackwell GPU Can Train AI Models with Trillions of Parameters

March 18, 2024

Nvidia's latest and fastest GPU, codenamed Blackwell, is here and will underpin the company's AI plans this year. The chip offers performance improvements from Read more…

Nvidia Showcases Quantum Cloud, Expanding Quantum Portfolio at GTC24

March 18, 2024

Nvidia’s barrage of quantum news at GTC24 this week includes new products, signature collaborations, and a new Nvidia Quantum Cloud for quantum developers. Wh Read more…

Houston We Have a Solution: Addressing the HPC and Tech Talent Gap

March 15, 2024

Generations of Houstonian teachers, counselors, and parents have either worked in the aerospace industry or know people who do - the prospect of entering the fi Read more…

Alibaba Shuts Down its Quantum Computing Effort

November 30, 2023

In case you missed it, China’s e-commerce giant Alibaba has shut down its quantum computing research effort. It’s not entirely clear what drove the change. Read more…

Nvidia H100: Are 550,000 GPUs Enough for This Year?

August 17, 2023

The GPU Squeeze continues to place a premium on Nvidia H100 GPUs. In a recent Financial Times article, Nvidia reports that it expects to ship 550,000 of its lat Read more…

Shutterstock 1285747942

AMD’s Horsepower-packed MI300X GPU Beats Nvidia’s Upcoming H200

December 7, 2023

AMD and Nvidia are locked in an AI performance battle – much like the gaming GPU performance clash the companies have waged for decades. AMD has claimed it Read more…

DoD Takes a Long View of Quantum Computing

December 19, 2023

Given the large sums tied to expensive weapon systems – think $100-million-plus per F-35 fighter – it’s easy to forget the U.S. Department of Defense is a Read more…

Synopsys Eats Ansys: Does HPC Get Indigestion?

February 8, 2024

Recently, it was announced that Synopsys is buying HPC tool developer Ansys. Started in Pittsburgh, Pa., in 1970 as Swanson Analysis Systems, Inc. (SASI) by John Swanson (and eventually renamed), Ansys serves the CAE (Computer Aided Engineering)/multiphysics engineering simulation market. Read more…

Choosing the Right GPU for LLM Inference and Training

December 11, 2023

Accelerating the training and inference processes of deep learning models is crucial for unleashing their true potential and NVIDIA GPUs have emerged as a game- Read more…

Intel’s Server and PC Chip Development Will Blur After 2025

January 15, 2024

Intel's dealing with much more than chip rivals breathing down its neck; it is simultaneously integrating a bevy of new technologies such as chiplets, artificia Read more…

Baidu Exits Quantum, Closely Following Alibaba’s Earlier Move

January 5, 2024

Reuters reported this week that Baidu, China’s giant e-commerce and services provider, is exiting the quantum computing development arena. Reuters reported � Read more…

Leading Solution Providers

Contributors

Comparing NVIDIA A100 and NVIDIA L40S: Which GPU is Ideal for AI and Graphics-Intensive Workloads?

October 30, 2023

With long lead times for the NVIDIA H100 and A100 GPUs, many organizations are looking at the new NVIDIA L40S GPU, which it’s a new GPU optimized for AI and g Read more…

Shutterstock 1179408610

Google Addresses the Mysteries of Its Hypercomputer 

December 28, 2023

When Google launched its Hypercomputer earlier this month (December 2023), the first reaction was, "Say what?" It turns out that the Hypercomputer is Google's t Read more…

AMD MI3000A

How AMD May Get Across the CUDA Moat

October 5, 2023

When discussing GenAI, the term "GPU" almost always enters the conversation and the topic often moves toward performance and access. Interestingly, the word "GPU" is assumed to mean "Nvidia" products. (As an aside, the popular Nvidia hardware used in GenAI are not technically... Read more…

Shutterstock 1606064203

Meta’s Zuckerberg Puts Its AI Future in the Hands of 600,000 GPUs

January 25, 2024

In under two minutes, Meta's CEO, Mark Zuckerberg, laid out the company's AI plans, which included a plan to build an artificial intelligence system with the eq Read more…

Google Introduces ‘Hypercomputer’ to Its AI Infrastructure

December 11, 2023

Google ran out of monikers to describe its new AI system released on December 7. Supercomputer perhaps wasn't an apt description, so it settled on Hypercomputer Read more…

China Is All In on a RISC-V Future

January 8, 2024

The state of RISC-V in China was discussed in a recent report released by the Jamestown Foundation, a Washington, D.C.-based think tank. The report, entitled "E Read more…

Intel Won’t Have a Xeon Max Chip with New Emerald Rapids CPU

December 14, 2023

As expected, Intel officially announced its 5th generation Xeon server chips codenamed Emerald Rapids at an event in New York City, where the focus was really o Read more…

IBM Quantum Summit: Two New QPUs, Upgraded Qiskit, 10-year Roadmap and More

December 4, 2023

IBM kicks off its annual Quantum Summit today and will announce a broad range of advances including its much-anticipated 1121-qubit Condor QPU, a smaller 133-qu Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire