High performance computing workflows in many industries are limited by memory bandwidth requirements, which are not satisfied by standard architectures. If memory bandwidth is a bottleneck, the sustained performance of real scientific applications will be limited.

Workflows impacted by a lack of memory bandwidth include a very broad range of applications such as computational fluid dynamics, structural analysis, and many applications that model physical systems. Much of the code used in these workflows can be well-adapted to vector processing. Specifically, the underlying code structures exhibit inherent parallelism, often only obscured by the programming style and not by the underlying mathematics. The lack of architectures really addressing the so-called memory wall was contemplated for many years in the HPC-community.

With these issues in mind, NEC will soon introduce a new vector-architecture nick-named Aurora, a solution that uses its industry-leading vector processor technology tightly coupled with industry standard scalar technology.

Expanding the use of vector processing

Organizations today must find ways to speed the time to results. The best way to accomplish this is to accelerate modeling and simulation algorithms. This offers many benefits. Individual jobs run faster, freeing up systems to run other workloads. And individuals or groups can more quickly explore different possibilities by modifying parameters, running more variations of the same job. But this speedup should not come at the price of increased coding complexity.

While some workloads benefit from acceleration technology such as GPUs, this benefit can only be achieved as a result of a large porting effort. An alternative that can be used on broad classes of applications is vector processing. Vector processing can satisfy the underlying need for memory bandwidth and thus leads to higher sustained performance without the need to rephrase huge amounts of codes in terms of complicated coding paradigms.

Today there is a growing need to make use of vector processing in a wide variety of industries. However, in most companies require a change in mindset. In the past some organizations have been hesitant about moving to vector processing because they either did not want to change their code or they thought it would be too much work.

That is no longer the case. Vectorization is close to inevitable on every platform, and therefore codes are adapted anyway. And in any case, smart optimization techniques like cache-blocking on scalar platforms are more difficult to apply than vectorization, let alone to rewrite a code using OpenCL or CUDA.

What is needed is a solution that delivers the desired performance benefits while allaying these concerns. Additionally, the system should be efficient in terms of performance per dollar or Watt. And their usage should also enable the scientist to work effectively.

A technology partner that can help

While almost all contemporary architectures utilize single instruction, multiple data (SIMD) parallelism in some way, only one vendor, NEC, is offering an architecture with real vector registers, entities which provide data to functional units continuously for a sequence of cycles by just one instruction.

Building on its long history developing vector processing systems for the most demanding workloads, NEC’s new Aurora vector system is designed to accelerate traditional memory bandwidth-intensive HPC workloads, as well as other applications such as big data analytics.

The solution is not just an accelerator that speeds up a small portion of code, as would be the case when using GPUs. Instead, the full code will run on a so-called vector engine, an accompanying x86-based platform just acts as kind of a frontend and a development environment.

Three design concepts guided the development of Aurora. They include:

Industry leading memory bandwidth performance: The individual cores of the new vector architecture will be  quite fast and can run code more efficiently. Similar to NEC’s SX Series, the solution offers industry-leading memory bandwidth per processor, core, power, and price. The cores are tailored to memory-intensive scientific applications and big data analytics applications.

Ease of use: Aurora offers a dedicated vector hardware/software environment. Specifically,

NEC’s optimized vector processing hardware and software are combined with a de facto standard environment such as Xeon clusters. This allows a workflow to start on an x86 system, and differently from accelerators, the entire application is passed to the vector engine. The x86-system supports the vector engine like a frontend, taking away all workload that does not relate to the application, daemons, and administrative processes.

Flexibility afforded by a hybrid solution: Aurora offers closely aligned scalar and vector machines. They can be used in hybrid configurations to tackle every kind of application, providing the appropriate hardware for each kind of code of a workload or workflow. This capability can be used in a workflow, for example, that requires pre- and post-processing of data or in a simulation such as a climate simulation that involves ocean code that needs vector processing and some atmospheric chemistry code that runs better on a scalar system. Software integration will include a common parallel filesystem, common scheduling, and an MPI that allows organizations to use both scalar and vector nodes in one application.

Summary

Organizations in many industries currently run simulation, modeling, and big data analytics applications that are limited because of memory bandwidth. Vector processors can deliver the necessary performance, but a solution must also be easy to use offering a standard environment and use of familiar Xeon clusters. Additionally, since many workflows include a mix of code where some algorithms can benefit from vector processing and others run more efficiently on scalar processors, any solution must be flexible allowing for such hybrid operations.

These are all areas where NEC Aurora can help. Aurora is a next-generation product that is designed to expand the use of the technology from traditional HPC problems to include a broader class of memory bandwidth-intensive applications used in organizations today.

NEC plans to make Aurora available for different environments including a rack-mounted server and supercomputer models, and it will be offered in a tower server model as well, so every scientist can develop code without continuous access to a high-end system.

For more information about meeting the demands of your memory bandwidth-intensive applications and workloads, visit: http://uk.nec.com/en_GB/emea/products/hpc/index.html