May 26, 2014

Intel True Scale Fabric – Made for HPC

Unlike today’s solid but somewhat stodgy basic enterprise computing, HPC tends to engender an enthusiasm and a fascination with the underlying technology that is unique to its practitioners.

But there is a downside. An overemphasis on advanced feeds and speeds and the joys of parallelism can distract us from what is really key – what HPC can do for its users.

For example, take the development of InfiniBand and its direct descendant, Intel’s True Scale Fabric, an advanced interconnect based on the InfiniBand architecture.

The original InfiniBand design was created to speed up channel interconnects between processors and I/O devices for the data center marketplace. The design was to support very good I/O performance, especially in moving large blocks of data around the data center.

Fascinating technology. But when it came time to move InfiniBand into a new venue – the world of HPC – there were some problems with InfiniBand’s design due to its focus on the data center marketplace.

Joe Yaworski, Director of Marketing for Intel’s High performance Fabric Operations, points out that the InfiniBand architecture needed a big makeover to function in an HPC environment.

In particular, InfiniBand was poorly equipped to handle’s HPC’s primary protocol – the Message Passing Interface (MPI). MPI requires the movement of large numbers of very small messages to achieve its performance and scalability. Through a sequence of fortuitous acquisitions, including acquiring QLogic’s InfiniBand business several years ago, Intel was able to develop an optimized interface for MPI that supplanted InfiniBand’s Verbs and connection-based design and opened the door for an optimized high-speed interconnect solution for HPC.   Evidence of this success was recently shared in an audio podcast.

Enter True Scale Fabric

“True Scale Fabric was designed from the ground up for HPC,” comments Yaworski.

He says that the technology offers improved HPC performance at a competitive price point, especially for systems where outstanding performance across large node counts is a requirement.

True Scale Fabric includes a comprehensive line up of adapters for various server form factors, a full line of edge or top of rack fabric switches that start as little as 18 ports and go as high as 864 ports, and all the requisite management software.

The True Scale Fabric architecture provides:

  • Improved connectionless design – Designed around an implementation that minimizes state information on the adapter, this implementation provides low end-to-end latency, even at scale.
  • Increased MPI message rate support – Through an optimized interface library layer called PSM (Performance Scaled Messaging), which is designed to optimize the performance and scalability of MPI based applications.

How About the Users

That’s just a quick glimpse of some of the technology involved. But what about that all-important question – how does True Scale Fabric benefit Intel’s HPC customers?

Because the True Scale Fabric is a highly scalable interconnect design, it helps harness the power of every processor core to deliver high message rates for parallel workloads. An excellent example is the Department of Energy that has implemented the Intel fabric at its Tri-Labs – Los Alamos, Lawrence Livermore, and Sandia National Labs.

“By installing the True Scale Fabric, the three labs were able to scale their applications even more than they anticipated,” Yaworski reports. “These installations have been so successful that the labs have continued to order servers based on Intel processors and the True Scale Fabric.”

Other government agencies are taking full advantage of the technology to build the HPC infrastructure required to handle projects such as modeling nuclear weapons effects, designing advanced weapon’s systems, and simulating security threats.

In addition to government labs, a number of industry verticals are using the technology as well. In the energy field, many of the world’s largest oil exploration companies are using the Intel fabric to speed up their HPC infrastructure in order to improve their complex analyses of exploration and recovery.

In the world of academia, scientists and researchers are using the technology to solve complex problems that range from exploring the multi-physics associated with creating green products, to how climate change will impact coastal cities.

In the life sciences, the technology is being used to speed up the analysis of the massive big data involved in such fields as molecular modeling and genomics.

Manufacturing companies are increasingly relying on HPC-based modeling and simulation to design their products and improve their competiveness and profitability. True Scale Fabric helps the HPC clusters and supercomputers employed by automotive, aerospace, electronics and other manufacturing industry verticals achieve the performance necessary to create new products and bring them to market faster.

Yaworski was particularly taken with Saudi Aramco’s use of the True Scale Fabric technology. He reports that at the Intel booth at last year’s International Supercomputer Conference, Saudi Aramco demonstrated the power of high performance computing paired with advanced visualization. They deconstructed the detailed image of an Audi R-5 into its component parts and put it all back together again. “It was a terrific demonstration and really showcased what can be done with True Scale,” he says.

Moving Toward Exascale

Like the HPC industry in general and Intel’s product lines in particular, the True Scale Fabric technology is constantly evolving. In addition to meeting today’s HPC needs, out there on the horizon is exascale, the computational Holy Grail that Intel has pledged to support.

“Among the technologies needed to realize exascale class systems is CPU fabric integration,” Yaworksi states. “We need the power, performance, density, and reliability to support say, 150,000 to 200,000 servers with tens of millions of cores. Today’s HPC fabric technology is terrific, but we need to reexamine the requirements that will take us to exascale.

“Because interconnect is a key technology, we will continue to pioneer next generation fabrics,” he concludes. “And we’ll be supporting a new generation of users who will be accomplishing things with tomorrow’s HPC systems that we can’t even begin to imagine.”

Share This