Aggregating Clusters Through Virtualization; Virtual SMP Benefits

By Shai Fultheim

December 16, 2008

Aggregation is a virtualization technique that makes multiple physical systems appear to function as a single logical system. The building blocks for this approach are the same building blocks used in scale-out (clustering): industry-standard servers and high-speed interconnects. Aggregation fundamentally replaces the functionality of custom and proprietary chipsets with software, leveraging commodity interconnects such as InfiniBand. It utilizes only a tiny fraction of the system’s CPUs and RAM to provide chipset-level services without sacrificing system performance.

By running a single logical system, customers reduce the management costs associated with managing multiple cluster nodes, and take advantage of large contiguous memory and unified I/O architecture. To understand how aggregation works, we will first explain the architecture of a traditional SMP system, and then dive into the details of the aggregation approach.

Traditional Multi-Processor Systems

Traditional multi-processor systems run a single operating system (OS). The OS interacts with the system using a well-defined hardware interface, which provides the OS with predefined services to use and control the hardware. These interfaces may include hardware detection and probing, memory ordering semantics, I/O space access and interrupt delivery mechanisms.

Intel’s Multiprocessor Specification allows a single copy of an operating system to run on a single CPU system as well as on a multi-CPU system with up to 255 CPUs. It details a well-defined interface that allows the OS to know exactly how to probe the hardware to determine what kind of system is running underneath it, and then behaves accordingly. This interface also handles the coordination of the underlying system with the OS. For a traditional multi-processor system, such interface is implemented in a silicon chipset.

In addition to the hardware interface, the multi-processor system consists of CPUs, memory and I/O subsystems. These components are all connected together with a proprietary backplane, often implemented by a chipset. Examples of such chipsets and backplanes are Intel’s FSB (Front Side Bus), AMD’s HT (Hyper-Transport), SUN’s CrossBar, SGI’s NUMALINK and IBM’s XA.

The chipset and the proprietary backplane (system interconnect) are the elements where multi-processor systems differ the most from each other and where the major cost of a high-end multi-processor systems is derived. The system interconnect is expensive because the more processors that are added to a system, the more complex it becomes to connect them all together in a manner that ensures both coherency and performance. Traditional multi-processor systems require the creation of a custom chipset to implement the system interconnect to allow processor, memory and I/O communication.

Software Approach: Aggregation

Aggregation reduces the costs associated with custom chipsets and backplanes. It requires multiple high-volume, industry-standard x86 systems (processor speed and amount of memory across boards do not have to be the same). In addition, a high-speed, low-latency interconnect serves as backplane. InfiniBand infrastructure excels in this regard. The aggregation software is loaded on each system board below the operating system layer.

One System

Once loaded into the memory of each of the system boards, the virtualization software aggregates the compute, memory and I/O capabilities of each system and presents a unified virtual system to both the operating system and the applications running above the OS. The aggregation software uses a software-interception engine in the form of a Virtual Machine Monitor (VMM) to provide a uniform execution environment. The software also creates the required BIOS and Advanced Configuration and Power Interface (ACPI) environment to provide the OS (and the software stack above the OS) a coherent image of a single system.

Coherent Memory

The software maintains cache coherency between the individual boards using multiple advanced coherency algorithms. These complex algorithms operate concurrently on a per-block basis, based on real-time memory activity access patterns. Board local-memory is leveraged together with caching algorithms to minimize the effect of interconnect latencies.

Shared I/O

The virtualization software aggregates I/O resources across all boards into a unified PCI hierarchy and presents them as a common pool of I/O resources to the OS and the application. The OS is able to utilize all the system storage and networking controllers toward providing high-I/O system capabilities.

The Advantages of Aggregation

Ease of Use

Aggregation simplifies scale-deployments by having a single system to manage compared to the complexities involved with managing a cluster. A single system removes the need for cluster file systems, cluster interconnect provisioning, application provisioning and installation and updating multiple operating systems and applications. The use of one operating system instead of one per node results in a significant savings in time and money during installation, as well as in ongoing management costs.

Simplified I/O Architecture

I/O requirements for a scale-out model can be very complex and costly, involving networked storage with accompanying costs related to additional HBAs, and FC switch infrastructure. Aggregation technology consolidates each individual server’s network and storage interfaces. I/O resource consolidation reduces the number of drivers, HBAs, NICs, cables, and switch ports and all the associated maintenance overhead. The user needs fewer I/O devices to purchase, manage and service with increased availability, resiliency and runtime scalability of I/O resources.

Large Memory System

For workloads that require a large contiguous memory, customers have traditionally used the scale-up approach. Aggregation provides a cost-effective alternative to buying expensive and large proprietary shared-memory systems for such workloads. It enables an application requiring large amounts of memory to leverage the memory of multiple systems, and reduce the need to use a hard-drive for swap or scratch space. Application runtime can be dramatically reduced by running simulations with in-core solvers or by using memory instead of swap for large-memory footprint models.

Aggregation thus provides a cost-effective virtual x86 platform with a large shared memory that minimizes the physical infrastructure requirements and can run both distributed applications, as well as applications requiring a large memory footprint at optimal performance on the same physical infrastructure.

Compute-Intensive, Large Core-Count Requirements

For workloads that require a high core count coupled with shared memory, customers have traditionally used proprietary shared-memory systems. Aggregation provides a very cost effective x86 alternative to these expensive and proprietary RISC systems.
Aggregation technology combines memory-bandwidth across boards, as opposed to traditional SMP or NUMA architecture where memory bandwidth decreases as the machine scales. This enables solutions based on aggregation technology to show close-to-linear memory bandwidth scaling, thereby delivering excellent performance for threaded applications.

Improved Utilization

Even in large cluster deployments in data centers, it makes sense to deploy aggregation, since fewer larger nodes mean less cluster complexity and better utilization of the infrastructure due to reduced fragmentation of the resources. An example can be found in the financial services industry, where organizations need to run hundreds or thousands of simulations at once. A common deployment model will involve hundreds of servers, where each will execute a few simulations. In this example, each cluster node is running a single application at 80 percent utilization. By using aggregation to create fewer larger nodes, every four aggregated systems can run another copy of the application, leveraging the underutilized resources and driving an additional 25 percent utilization.

ScaleMP’s vSMP Foundation is an example of the aggregation concept. The company has taken an approach that makes a tradeoff between memory latency and memory bandwidth.

Memory Bandwidth vs. Memory Latency

vSMP Foundation uses caching technologies to provide parallel access to system memory. With vSMP Foundation, data migration and replication are employed to maximize system memory bandwidth. The additional system memory bandwidth is used to mask the backplane latencies. ScaleMP uses standard interconnect technology versus the custom built back planes used in traditional multi-processors. While the backplane latency of the InfiniBand interconnect is higher than traditional multi-processor backplanes, the additional memory bandwidth offsets this higher latency. One of the keys to appreciating vSMP Foundation’s ability to mask backplane latency and provide superior performance is the understanding of the fundamentals behind efficiency in memory management. At its elemental level, efficiency can be defined as:

Efficiency = 1 – (Access x Latency)

Where:

Access represents the number of times a processor has to reach out to memory that is not within the processor cache (i.e., on main memory, requiring access via the backplane).

Latency represents the amount of processor wait time such memory requires each time it is accessed.

The efficiency of a system can be improved by reducing the number of times the processor accesses the backplane or reducing the latency of each access to the backplane. Typically, the access is defined by the nature of the application, and latency is based on the technology of the backplane.

Historically, the industry has improved performance by focusing significant R&D on reducing latency in each new generation of products (backplanes, memory-speed, etc.). The assumption was that the access patterns were driven by the applications and, hence, largely out of the control of the system vendors.

ScaleMP’s basic approach was to replace proprietary chipsets and backplanes with InfiniBand interconnects and reduce the number of times a processor has to access the backplane for memory operations on another physical board.

Non-Uniformed Memory Architecture (NUMA) is a common multi-processor architecture which inherently support system scaling by adding additional system nodes. NUMA drawbacks are result of non-homogenous memory access latency, which require operating system and application awareness. ScaleMP utilizes a combination of NUMA and Cache Only Memory Architecture (COMA) in conjunction with a massive cache (typically 5-10 percent of the system’s RAM), to trade off the backplane latency with the use of redundant RAM for caching. The backplane latency is mitigated using software-driven adaptive caching techniques and achieving better systems economics by leveraging commodity memory costs versus proprietary backplanes and chipsets.

ScaleMP utilizes multiple memory coherency algorithms that are selected based on several aspects of the application behaviors, such as historical memory access pattern, on-the-fly code analysis and I/O behavioral analysis. In essence, in spite of having higher backplane latency versus traditional multi-processor systems, vSMP Foundation techniques for memory access reduction are designed to offset the disadvantage of higher-latency, commodity industry-standard interconnects. With the progress that InfiniBand is making in improving latency from SDR to DDR to QDR, the latency gap between an InfiniBand interconnect and a proprietary interconnect is shrinking.

By leveraging industry-standard processors and systems coupled with industry-standard interconnects, aggregation creates a new paradigm for high performance computing and represents a step toward delivering better performance with lower cost and less complexity. Aggregation overcomes the fundamental limitation of clusters for applications that require large shared memory, and it addresses the barriers many technical computing customers have when it comes to adopting clusters, namely a lack of appropriate IT skills to install and manage them. The drawbacks associated with traditional SMP systems, such as high cost and vendor lock-in, are likewise avoided.

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