ESnet Launches Architecture to Help Researchers Deliver on Data-Intensive Science

By Jon Bashor

April 26, 2012

The U.S. Department of Energy’s Energy Sciences Network, or ESnet, provides reliable high-bandwidth network services to thousands of researchers tackling some of the most pressing scientific and engineering problems, such as finding new sources of clean energy, increasing energy efficiency, understanding climate change, developing new materials for industry and discovering the nature of our universe. To support these research endeavors, ESnet connects scientists at more than 40 DOE sites with experimental and computing facilities in the U.S. and abroad, as well as with their collaborators around the world. ESnet is managed for DOE’s Office of Science by Lawrence Berkeley National Laboratory.

As science becomes increasingly data-intensive, the ESnet staff regularly meets with scientists to better understand their future networking needs, then develops and deploys the infrastructure and services to address those requirements before they become a reality. One example of this is the Advanced Networking Initiative, a prototype 100 gigabits-per-second networking connecting the DOE Office of Science’s top supercomputing centers in California, Illinois and Tennessee, and an international peering point in New York. This 100 Gbps prototype is now being transitioned to production and will be rolled out to all other connected DOE sites in the coming year.

In order to help these research institutions fully capitalize on this growing availability of bandwidth to manage their growing data sets, ESnet is now working with the scientific community to encourage the use of a network design model called the “Science DMZ.” The Science DMZ is a specially designed local networking infrastructure aimed at speeding the delivery of scientific data. In March 2012, the National Science Foundation supported the concept by issuing a solicitation for proposals from universities to develop Science DMZs as they upgrade their local network infrastructures.

Leading the development of the Science DMZ effort at ESnet is Eli Dart, a network engineer with previous experience at Sandia National Laboratories and the National Energy Research Scientific Computing Center. In this interview conducted by Jon Bashor of Berkeley Lab, Dart answers some basic questions about the nature of the project and its principle goals.

Jon Bashor: What is the Science DMZ and where did the Science DMZ idea come from?

Eli Dart: In its purest form, it’s an element of the overall network architecture, typically a dedicated portion of a site or campus network, located as close to the network perimeter as possible, that serves only high-performance science applications. The intent of the Science DMZ is to simplify the deployment and support of high-performance and data-intensive science applications that rely on high-speed networking for success. These applications have unique network requirements that typically cannot be met by networks that are optimized for normal business operations like web browsing, procurement and financial systems, and the like. The idea itself came from two places.

The concept of a DMZ network originated in the network security space where so-called network “demilitarized zones” or DMZs are used to provide a dedicated portion of the network near the site perimeter specifically configured to support services that interact with the outside world. These services often include authoritative DNS, incoming email, outward facing websites, etc. These services usually fall under a security policy that’s different than the one for the rest of the enterprise architecture.

You can extend that notion to build a dedicated piece of the network specifically for high performance scientific applications, again located at or near the perimeter, and with hardware you know can handle these applications. The Science DMZ is not configured to handle the standard enterprise or business functions, such as email and web servers, desktop applications, and so forth. These typically need a massive security infrastructure to protect them, and the security measures required to protect business servers and desktop applications typically cause problems for high-performance applications. The Science DMZ model explicitly separates the science traffic from general-purpose network traffic, and allows appropriate security policies and enforcement mechanisms to be applied to each.

The second source for the Science DMZ concept came from working with TCP, or the Transmission Control Protocol. While most science applications that need reliable data delivery use TCP-based tools for data movement, TCP’s interpretation of packet loss can cause performance issues. TCP interprets packet loss as network congestion, and so when loss is encountered TCP dramatically reduces its sending rate – slowing the data transfer. In practice even a tiny amount of loss (much less than 1%) can be enough to reduce TCP performance by over a factor of 100.

For years people have been trying to fix TCP (with some success), but packet loss combined with high latency is a serious performance killer. It’s easier to build an infrastructure to provide loss-free IP service and to accommodate TCP rather than change it – this is what the Science DMZ model aims to accomplish.

Bashor: What makes up the Science DMZ model?

Dart: The Science DMZ itself is a portion of the network, at or near the site perimeter, which is specifically configured to support high-performance science applications. There are several key aspects to the Science DMZ.

First, it must be built with capable equipment that can handle high-rate flows without dropping packets. Typically, that means good equipment (not cheap wiring closet switches) with enough output buffer space to handle bursty high-rate long-distance TCP flows. The switches and routers need to be able to accurately account for packets (especially the ones they drop) so that packet loss can be accounted for and its cause fixed.

Second, data transfer should be done on dedicated servers – Data Transfer Nodes, or DTNs – that are designed and configured for the purpose. Their TCP stacks need to be tuned and they need access to high-speed storage. We have seen successful DTN implementations using high-speed local RAID as well as GPFS or Lustre filesystems, the parallel filesystem model is typically found at supercomputer centers.

Third, a Science DMZ needs test and measurement infrastructure, typically perfSONAR that allows you to identify any issue that may be causing performance issues. Many problems that are real performance killers are what we call “soft failures.” A soft failure causes performance degradation so that the network is not useful for data-intensive science but does not cause an outage that identifies the failing component. The only way to find these is to independently test the infrastructure to locate the problem – if perfSONAR is already deployed, this is much easier than if the first step of the process is to find and deploy a test machine and the second step is to get the site at the other end to find a spare box and deploy it.

Finally, the Science DMZ incorporates a security policy that is tailored to the science applications rather than to general-purpose business computing. You don’t need to scan 50TB of simulation output for email viruses, and you don’t run an email client on your Data Transfer Node. So, why conflate the security policies and enforcement mechanisms for the two, especially when doing so will effectively compromise the science mission? Firewalls and other security enforcement boxes are typically unable to handle the throughput needed for data-intensive science – and they essentially never support advanced science services such as virtual circuits or software-defined networking.

Bashor: Why does it matter?

Dart: The real reason all this matters is that the current and future generations of scientific instruments are producing data at a level we’ve never seen before. Based on our projections, ESnet is expected to carry over 100 petabytes of data per month by 2015. And there is the potential for stupendous scientific advancements in that data deluge. The challenge is to figure out how to get the science done without spending the bulk of your time doing data management. Scientists are physicists, chemists, biologists, geneticists and so on, but they are seldom network experts. They are scientists.

The data volumes are becoming large enough that the systems and networks are not capable of handling them if the equipment is configured to default settings or to accommodate business applications. There’s a need for an infrastructure that supports data-intensive science. That infrastructure needs to be designed for data mobility, which means you can get the data where you need it, when you need it. In some cases, the analysis code is on a system close to the data, while other times the scientist wants to analyze the data on local resources – we need to support it all. Data-intensive science is what we’re all going to be doing for the next decade or more.

Bashor: Can you describe a typical user who would benefit from having a Science DMZ?

Dart: The main benefit of the Science DMZ is that the scientist who needs to move data doesn’t have to first troubleshoot the infrastructure in order to use it. Scientists should not have to fix the network, the data transfer servers, and so forth before they can get to work.

There really isn’t a typical user, but there are some basic commonalities. One example could be data taken from a beamline at DOE’s Advanced Light Source. A data transfer node has been set up and Globus Online installed for users who need to fetch the data. Then you have the well-known Large Hadron Collider, which has several primary Tier 1 centers feeding data to the Tier 2 centers. This requires significantly more infrastructure. In both cases, you need to make sure the network is designed correctly so that data transfer tools work correctly. These fundamental principles benefit all users.

Bashor: How does ESnet play into this equation?

Dart: ESnet is the high-performance network for DOE’s Office of Science. It’s the backbone network infrastructure for the national laboratory system, supporting science at those labs. Through our 25 years of experience serving the scientific community, we have become a central repository for the expertise to support high-performance networking. So, part of our job is to be available to support scientists at the labs and their collaborators, such as researchers at universities.

The assumption is that the high-performance network infrastructure exists to support all parts of these modern scientific collaborations. The services must be consistent from end to end – from scientist to scientist – now matter where they may located and regardless of who owns the pieces of the infrastructure. For example, if scientists at the SLAC Linear Accelerator Center are sharing data with colleagues at a Max Planck Institute in Germany, the data moves from SLAC’s local network over ESnet to GEANT, the pan-European research network, then over Germany’s DFN network and onto the local network at the institute – crossing five different domains, owned and operated by five different organizations. ESnet has built partnerships with the global ecosystem of research and education networks so that if a network problem occurs, we can work collaboratively to quickly resolve it – wherever it is.

Bashor: The NSF recently cited Science DMZ as an upgrade that universities should consider as they work to enhance their overall IT infrastructure. Your thoughts on this?

Dart: We think it’s wonderful. The infrastructure that will be built with those funds will enable discoveries that otherwise would not be possible. It’s a critical investment in the scientific infrastructure of this country.

As I said, we’re all going spend the next decade or more supporting data-intensive science, so we need to get the infrastructure right. It needs to be adaptable, flexible and expandable. We can see what’s coming in the next one to three years. In some fields, the cost of generating data has fallen to almost zero. In genome sequencing, the cost per genome has fallen off a cliff. The cost of a raw megabyte of DNA sequence is now less than 10 cents. In July 2001, it was about $4,500. What this means is that we are entering a world where scientific productivity is gated on data analysis, not data generation.

In physics, new detectors will capture data in the terabyte-per-second range, with data analysis and reduction built into the detectors, so that only the data the researchers are really interested in will be kept. This is already happening at the LHC. The ATLAS detector generates about a petabyte of data a second. It’s sent through a multi-stage trigger farm where it’s reduced to about 2.5 gigabits per second coming out. Now many other science domains are getting into this same situation.

Looking 10 years out is beyond the current planning and budget outlooks – and well outside the scope of a single procurement or a single technology. This puts the work into the architecture space, not the technology or device space. We do know that everything about the data is growing exponentially, but not the funding. So we need to design a system that works well in general and is adaptable.

If you want to do capability-class science, you need to have capability-class infrastructure. You have to have the resources appropriate to get the most return on your scientific investment.

Bashor: ESnet has a number of projects to improve end-to-end network performance through testing and measurement. Can you talk about those briefly?

Dart: Performance testing and measurement is absolutely critical. If we go back to the need to accommodate TCP because packet loss is the number one enemy of data-intensive science, we have to be able to find and fix any problems quickly. Because issues can arise anywhere on the network path which can include multiple administrative domains, you need to have the means to individually test the paths, and take out or reconfigure the problem areas.

For this reason, ESnet – with Internet2 and several other collaborators – helped develop perfSONAR, an infrastructure for network performance monitoring, making it easier to solve end-to-end performance problems on paths crossing several networks. ESnet has test and measurement capabilities at every hub site and router on our network. You have to have this infrastructure in place before a problem occurs – this allows you to find and fix the problem in hours or days, not months.

Another service for improving end-to-end performance is OSCARS, ESnet’s On-Demand Secure Circuits and Advance Reservation System. OSCARS provides multi-domain, high-bandwidth virtual circuits that guarantee end-to-end network data transfer performance. With a Science DMZ, OSCARS can touch down at an institution, along with other science-specific services. This allows for capability-class services to be used without interfering with the enterprise system. The bottom line is that science opportunities have a better chance of not being missed.

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!

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion XL — were added to the benchmark suite as MLPerf continues 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 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…

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion 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…

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