The 3D Torus architecture and the Eurotech approach

By Nicole Hemsoth

June 20, 2011

The 3D Torus architecture and the Eurotech approach

The ability of supercomputers to progressively run job faster meets the computational needs of both scientific research and an increasingly higher number of industry sectors.

Processor power is centerpiece to determine the performance of an HPC system, but it is not the only factor. One of the key aspects of parallel computers is the communication network that interconnects the computing nodes. The network is the one guaranteeing fast interactions between CPUs and allowing the processors to cooperate: this is essential to solve complex computational problems in a fast and efficient way.

Together with speed, HPC systems are increasingly asked to be more available. Downtime can affect a high performance machines quite badly. It appears clear that a reliable average machine with a great uptime is better than a high performance one with a low MTBF (mean time before faults): ultimately, the former would process more jobs in a week than the latter.

One additional challenge with large systems is scalability, so the ability to add nodes to a cluster without affecting performance and reliability or affecting them as little as possible. Petascale and then exascale installations require and will require hundreds of thousands of cores to efficiently work together.

It is also paramount for future machines to consume less energy, being cost and availability of electrical power an issue that is becoming the most demanding challenge to overcome along the road to exascale computing.

3D torus connectivity

In a computer cluster, the way the nodes are connected together could provide some determinant help to solve the above mentioned issues.

Despite being available for quite a while, the torus architecture has now the potential to surge from niche application to mainstream. This is because, like never before, we face nowadays some severe challenges posed by a raising number of nodes. The problem, before being one of performance, is one of topology and scalability. The more a system grows, the more fat tree switched topologies show limits of cost, maintainability, consumption, reliability and, above all, scalability.

Connecting nodes using a 3D Torus configuration means than each node in a cluster is connected to the adjacent ones via short cabling. The signal is routed directly from one node to the other with no need of switches. 3D means that the communication takes places in 6 different “directions”: X+, X-, Y+, Y-, Z+, Z-. In practical terms, each node can be connected to 6 other nodes: in this way, the graph of the connections resembles a tri-dimensional matrix.

Such configuration allows the addition of nodes to a system without degrading performance. Each new node is joint as an addition of a grid, linked to it with no extensive cabling or switching. While scaling linearly, with little or no performance loss, is strictly true for those problems that heavily rely on next neighbor communication, it is also true that, avoiding switches, hundreds of nodes can be added without causing problems of clogged links or busy fat tree switch leafs. This comes without considering that the addition of a node in a large system happens with much less working and potential troubles on a 3D torus network than on switched fat tree one. 

The pairwise connectivity between nearest neighbor nodes of a 3D Torus configuration helps to reduce latency and the typical bottlenecks of switched networks. Being the connections between nodes short and direct, the latency of the links is very low: this affects positively the machine performance, especially for solving local patters problems, which can be effectively mapped onto the matrix mentioned above. The switchless nature of the 3D Torus facilitates fast communication between nodes.

Switches are also potential points of failure. Decreasing their number should improve the operational functioning of a system. In other words, the 3D Torus makes a system more agile, so more prompt to react to failures: if a connection or a node fails, the affected communication can be routed in many different directions. The inherent nature of the 3D Torus is the connectivity of each node to its nearest neighbors to form a tridimensional lattice that guarantees multiple ways for a node to reach another one.

Eliminating costly and power-hungry external spine and leaf switches, as well as their accompanying rack chassis and cooling systems, torus architectures also contribute to reduce installation costs and energy consumption.

Applications

When it comes to applications that can fully benefit from the 3D Torus configuration we could touch one of the caveats of this intelligent connection schema.

The maximization of the performance of the 3D Torus takes place with a subset of problems which is rather large but specific. These are local pattern problems, which typically deal with modeling systems whose functioning/reaction depends on adjacent systems. Typical examples are computer simulations of Lattice QCD and fluid-dynamics. More in general, many Monte Carlo simulations and embarrassingly parallel problems can exploit the full performance advantage of the 3D Torus architecture, making the range of possible applications quite vast, especially within the field of scientific research.

Problems that require all to all dialogue between nodes are less prone to exploit the full performance of the 3D torus interconnection. However, independently from the type of application and problem, the 3D torus still bears the massive advantage of scalability and serviceability, contributing also to increase the availability of the systems and reduce power consumption. In case of large systems, it may be so advantageous to resort to the 3D Torus architecture that the perfect match with the problem that better maximize the computational performance may well become secondary.

The Eurotech approach

It is rather interesting to analyze what Eurotech, a leading European computer manufacturer, has done with the 3D Torus network of their supercomputer line Aurora.

Eurotech wanted to leverage the 3D Torus benefits for their high end products, but at the same time leave to the users the flexibility and the freedom to run all the applications they need.

Taking in account these diverging characteristics, Eurotech and its scientific partners took and approach called Unified Network Architecture in designing the Aurora datacenter clusters. This fundamentally means that the Aurora systems have 3 different networks working in concomitance on the same machine: 2 fast independent networks (Infiniband, 3D Torus) and a multi-level synchronization network.

The coexistence of Infiniband and 3D Torus facilitates flexibility of use: depending on the application, one or the other network can be utilized. The synchronization networks act at different levels, synchronizing the CPUs and de facto reducing or eliminating the OS jitter and hence making the system more scalable.

Torus topologies have traditionally been implemented with proprietary, costly application-specific integrated circuit (ASIC) technology. Eurotech chose to drive the torus with FPGAs, injecting more flexibility in the hardware, and to rely both on a GPL and on a commercial distribution for the 3D torus software. The 3D torus network is managed by a network processor implemented in the FPGAs, which interfaces the system hub through two x8 PCI Express Gen 3 connections, for a total internal bandwidth of 120Gbs.

Each link in the torus architecture is physically implemented by two lines (main and redundant) that can be selected (in software) to configure the machine partitioning (full 3D Torus or one of the many 3D sub-tori available). In this way, redundant channels allow system repartitioning on-the-fly. The possibility of partitioning the system into sub-domains permits to create system partitions that communicate on independent tori, effectively creating different execution domains. In addition, each subdomain can benefit from a dedicated synchronization network.

As an example of partitioning, if a backplane with 16 boards is considered, the available topologies for the partitioning of the machine in smaller sub-machines with periodic boundaries (sub-tori) are:

Half Unit: 2 x [1:2*NC] x [1:2*NR]

Unit: 4 x [1:2*NC] x [1:2*NR]

Double Unit:8 x [1:2*NC] x [1:2*NR], Rack: 8 x 2*NC x 2, Machine: 8 x 2*NC x 2*NR

Chassis: 16 x [1:NC] x [1, 2,  2*NR], Rack: 16 x NC x 2, Machine: 16 x NC x 2*NR

Where

–         NC : Number of chassis in a rack (8).
–         NR : Number of racks in a machine (from 1 to many hundreds)

Partitioning, FPGAs, redundant channels, synchronization networks are some of the unique characteristics that Eurotech wanted in the torus architecture to create Intel based clusters with the flavor of a special machines.

Eurotech

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