Advancing Modular Supercomputing with DEEP and DEEP-ER Architectures

By Sean Thielen

February 24, 2017

Editor’s note: In this contributed feature, Sean Thielen details the innovative architectural concepts coming out of the DEEP and DEEP-ER projects, and follows the evolution of the Cluster-Booster architecture, a heterogeneous hardware design developed at the Jülich Supercomputing Centre.

Knowing that the jump to exascale will require novel architectural approaches capable of delivering dramatic efficiency and performance gains, researchers around the world are hard at work on next-generation HPC systems.

In Europe, the DEEP project has successfully built a next-generation heterogeneous architecture based on an innovative “cluster-booster” approach. The new architecture can dynamically assign individual code parts in a simulation to different hardware components based on which component can deliver the highest computational efficiency. It also provides a foundation for a modular type of supercomputing where a variety of top-level system components, such as a memory module or a data analytics module for example, could be swapped in and out based on workload characteristics. Recently, Norbert Eicker, head of the Cluster Computing research group at Jülich Supercomputing Centre (JSC), explained how the DEEP and DEEP-ER projects are advancing the idea of “modular supercomputing” in pursuit of exascale performance.

Why go DEEP?

Eicker says that the use of vectorization or multi-core processors have become the two main strategies for acceleration. He noted that the main advantages in general purpose multi-core processors include high single-thread performance due to relatively high frequency along with their ability to do out-of-order processing. Their downsides include limited energy efficiency and a higher cost per FLOP. Accelerators, such as the Intel Xeon Phi coprocessor or GPUs, on the other hand are more energy efficient but harder to program.

Given the different characteristics of general purpose processors and accelerators, it was only a matter for time before researchers began looking for ways to integrate different types of compute modules into an overall HPC system. Eicker said that most efforts have involved building heterogeneous clusters wherein standard cluster nodes are connected using a fabric and then accelerators are attached to each cluster node.

Figure 1: An example of a basic architecture for a heterogeneous cluster.

Per Eicker, this heterogeneous approach has drawbacks, including the need for static assignment of accelerators to CPUs. Since some applications benefit greatly from accelerators and others not at all, getting the ratio of CPUs to accelerators right is tricky and inevitably leads to inefficiencies. Eicker explained that the idea behind the DEEP project was to combine compute resources into a common fabric and make the accelerating resources more autonomous. The goal was to not only enable dynamic assignments between cluster nodes and the accelerator, but also to enable the accelerators to run a kind of MPI so the system could offload more complex kernels to the accelerators rather than needing to always rely on the CPU.

The building blocks of a successful prototype

Work on the prototype Dynamical Exascale Entry Platform (DEEP) system began in 2011, and was mostly finalized toward the end of 2015. It took the combined efforts of 20 partners to complete the European Commission funded project. The 500 TFLOP/s DEEP prototype system includes a “cluster” component with general-purpose Intel Xeon processors and a “booster” component with Intel Xeon Phi coprocessors along with a software stack capable of dynamically separating code parts in a simulation based on concurrency levels and sending them to the appropriate hardware component. The University of Heidelberg developed the fabric, which has been commercialized by EXTOLL and dubbed the EXTOLL 3D Torus Network.

Figure 2: The DEEP cluster-booster hardware architecture. The cluster is based on an Aurora HPC system from Eurotech. The booster includes 384 Intel Xeon Phi processors interconnected by Extoll fabric.

Given the unusual architecture, the project team knew it would need to modify and test applications from a variety of HPC fields on the DEEP system to prove its viability. The team analyzed each selected application to determine which parts would run better on the cluster and which would run better on the booster, and modified the applications accordingly. One example is a climate application from Cyprus Institute. The standard climate model part of the application runs on the cluster side while an atmospheric chemical simulation runs on the booster side, with both sides interacting with each other from time to time to exchange data.

The new software architecture

One of the most important developments of the DEEP project is a software architecture that includes new communication protocols for transferring data between network technologies, programming model extensions and other important advancements.

Figure 3: The DEEP software architecture includes standard software stack components along with some new components developed specifically for the project.

While left- and right-hand sides of the architecture in figure 3 are identical to the standard MPI-based software-stacks of most present day HPC architectures, the components in the middle add some important new capabilities. Eicker explained that in the DEEP software architecture, the main part of applications and less scalable code are only run on the cluster nodes and everything starts on the cluster side. What’s different is that the cluster part of the application can collectively start a crowd of MPI-processes on the right-hand side using a global MPI.

The spawn for the booster is a collective operation of cluster processes that creates an inter-communicator containing all parents on one side and all children on the other. For example, the MPI_COMM_WORLD or a subset of processes on the cluster side, collectively called the MPI_Comm_spawn function, can create a new MPI_COMM WORLD on the booster side that is capable of standard MPI communication. Once started, the processes on the booster side can communicate amongst each other and exchange messages, making it possible to offload complex kernels to the booster.

Using MPI to bridge between the different fabrics in the cluster and booster may seem like it would significantly complicate the lives of application developers. However, Barcelona Supercomputing Center invented what is basically a source-to-source compiler, called the OmpSs Offload Abstraction compiler that does much of the work. Developers see a familiar looking cluster side with an Infiniband-based MPI and a booster side with an EXTOLL-based MPI. Their job is to annotate the code to tell the compiler which parts should run on the cluster versus the booster. The OmpSs compiler introduces the MPI_Comm_spawn call and the other required communication calls for sharing data between the two code parts.

Eicker explained that the flexible DEEP approach has many advantages, including options for multiple operational modes that enable much more efficient use of system resources. Beyond the specialized symmetric mode described above, the booster can be used discretely, or as a pool of accelerators. He used applications that could scale on the Blue Gene system as an example, noting they be run entirely on the booster side with no cluster interaction.

From DEEP to DEEP-ER

Plans for the DEEP-ER (Dynamical Exascale Entry Platform – Extended Reach) phase include updating the booster to include the latest generation of Intel Xeon Phi processors. The team is also exploring how on-node Non-Volatile Memory (NVM), network attached memory and a simplified interface can improve the overall system capabilities.

Figure 4: The DEEP-ER cluster-booster hardware architecture.

Eicker said that since Xeon Phi processors are self-booting, the upgrade will make the hardware implementation easier. The team also significantly simplified the interface by using the EXTOLL fabric throughout the entire system. The global use of the EXTOLL fabric enabled the team to eliminate the booster interface nodes and the DEEP cluster-booster protocol. The DEEP-ER system will use a standard EXTOLL protocol running the two types of nodes. The EXTOLL interconnect also enables the system to take advantage of the network attached memory.

One of the main objectives of the DEEP-ER project is to explore scalable I/O. To that end, the project team is investigating the integration of different storage types, starting from the disks using NVM while also making use of the network attached memory. Eicker said the team is using the BeeGFS file system and extensions that enable smart caching to local NVMe devices in the common namespace of the file system to help improve performance as well as SIONlib, a scalable I/O library developed by JSC for parallel access to task-local files, to enable more efficient local tasking of I/O. Exascale10 I/O software from Seagate also sits on top of the BeeGFS file system, enabling the MPI I/O to make use of the file system cache extensions.

Beyond I/O, the DEEP-ER project is also exploring how to improve resiliency. Eicker noted that because the offloaded parts of programs are stateless in the DEEP approach, it’s possible to improve the overall resiliency of the software and make functions like checkpoint restart a lot more efficient than standard approaches.

Toward modular supercomputing

Each phase of the DEEP project is an important step forward toward modular supercomputing. Eicker said that the DEEP cluster-booster concept showed that it’s possible to integrate heterogeneous systems in new ways. With DEEP-ER, the combination of the NAM and network attached storage add what is essentially a memory booster module. Moving forward, there are all kinds of possibilities for new modules, according to Eicker. He mentioned an analytics module that might look like a cluster, but include more memory or different types of processors, or a module that acts as a graphics cluster for online visualization.

Figure 5: The end goal of the DEEP project is to create a truly modular supercomputer, which could pave the way for increasingly specialized modules for solving different types of supercomputing challenges.

The ultimate goal of the DEEP project is to build a flexible modular supercomputer that allows users to organize applications for efficient use of the various system modules. Eicker said that the DEEP-ER team hopes to extend its JURECA cluster with the next-generation Xeon Phi processor-based booster. Then the team will begin exploring new possibilities for the system, which could include adding new modules, such as a graphics, storage and data analytics modules. The next steps could even include a collaboration with the Human Brain Project on neuromorphic computing. And these ideas are only the beginning. The DEEP approach could enable scientists to dream up new modules for tackling their specific challenges. Eicker acknowledges that there is much work to be done, but he believes the co-design approach used by the DEEP team will continue to drive significant steps forward.

Watch a short video capturing highlights of Eicker’s presentation.

About the Author

Sean Thielen, the founder and owner of Sprocket Copy, is a freelance writer from Portland, Oregon who specializes in high-tech subject matter.

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