Home > HPC > Meeting Research Needs in a Changing HPC Landscape

Meeting Research Needs in a Changing HPC Landscape

A look at adapting environments through hybrid, clustered, open IT architectures

Ken Edgecombe, Ph.D.

As business pressures rise, high performance computing has continued to transition into the mainstream and is fast emerging as an attractive means for organizations to do more with less. Businesses are becoming increasingly reliant on technology innovation to deliver both customer and shareholder value, while academic research organizations are under equally severe pressure to fuel the latest and greatest applications that will, over time, impact the way the world operates. Research institutions and businesses alike must work to dissolve the barriers that have traditionally slowed time-to-market and to ensure the correct tools and infrastructure are in place to achieve maximum business results.

Challenges in the field
Despite many governments and businesses praising the benefits of HPC, key challenges have emerged around perceived lack of value and associated IT costs, efficiency, performance, power and reliability. The complexities around system architectures are a specific area that research organizations and academia must overcome before the commercial world can follow their example. What also is apparent is the need for adaptable, easy-to-deploy systems that simplify workloads and offer greater scalability and flexibility to meet changing demands. Given these growing pressures, many are reevaluating their IT architecture and asking themselves what combination of commodity and complex hardware is the optimum server choice to get ahead in HPC.

The server architecture conundrum
Over the last five to seven years, clusters of x86 systems have come to dominate the supercomputing landscape, once known for its high-end SMP servers and mainframes. At the High Performance Computing Virtual Laboratory (HPCVL), we’ve experienced the pendulum swinging back to support both a hybrid of chip multithreading (CMT) clusters and high-end SMP servers, showing a trend toward multithreaded architectures in HPC. That being said, there’s no doubt that commodity x86 architectures have proved popular due to ease-of-use, low cost and high-level processing capabilities for applications, including rendering in the media and entertainment space and pattern matching and searching in scientific fields.

However, in recent years, power and memory demands have grown beyond expectations, particularly in relation to chord size and calculation at certain stages of the programming cycle. Chip temperatures, latencies and CPU development have meant that certain speeds and feeds are not providing enough throughputs and, put simply, are not keeping up — an area where both CMT and SMP systems can offer significant power and memory gains.

The challenge HPC developers and architects face today is increasingly complex data to access, manage and deploy in an intricate IT battleground where power, speed and reliability needs are becoming increasingly demanding. Subsequently, the role of a multithreaded environment is becoming more defined, with market predictions by IDC implying that the UNIX-based CMT systems market is poised for strong growth.

So, what has taken both CMT clusters and high-end SMP multithreaded systems so long to proliferate in the HPC space? Aside from a generally higher cost per box and more complex architecture to manipulate, one unresolved pain point is the fact that few software vendors have embraced the multithreaded evolution. Current licensing models lack complexity and are limited by process or CPU evaluation, meaning that, although a user has typically slower threads overall, it can in fact equate to more licenses and increased cost.

This is a community and industry issue as opposed to a vendor issue, which will no doubt undergo much more debate. Until resolution is found, regardless of specific product deployment, users should look to a software platform that is dynamic, open and ‘thread safe’ in its characteristics. At HPCVL, our researchers also have favored platforms that include dynamic tracing features.

Do high end systems have a role in HPC environments in the future? From my point of view, absolutely yes — and here’s why. Despite the questions that remain around pricing structures, it’s becoming evident that, as an industry, we are entering an increasingly multithreaded world. As such, there are some very real benefits to using a hybrid of both high-end SMP and CMT clustered systems in HPC environments. CMT clusters can offer some strong value propositions in terms of throughput. By using relatively low-performing but multiple cores, each with multiple threads on a single chip, it’s possible to effectively scale and increase the memory footprint, on the assumption that single threads that "do less" may still provide for higher performance if, collectively, the multiple threads can be utilized to get problems solved very quickly. This is a particularly attractive platform for HPC developers working on parallel algorithms and also can provide users significant cost gains given overall thread vs. hardware cost. This rationale may come as a surprise to many, perhaps due in part to the single-process school of thought to which we have subconsciously become accustomed.

Vendors aside, the following characteristics should high priority when choosing the right HPC server architecture:
• Which applications are considered critical to your business success? Choose an architecture that is open, flexible and able to scale easily to deal with sudden unpredicted growth
• How are users going to take full advantage of the new technology? The right IT environment needs to be facilitated by extensive support and services so that users can work in a sufficiently self-serving manner, requiring minimal guidance and intervention
• What are your access requirements? With working environments becoming increasingly distributed, it’s important to consider remote access that is reliable, secure and able to cope with varying power demands

HPCVL: A hybrid of high-end SMP and CMT clusters 
After considering the pros and cons of the different HPC server architectures available, you may find that the ideal solution for your business needs is, in fact, a hybrid of architectures. Two years ago at HPCVL, we had issues associated with system performance, user capacity, energy consumption and data center costs. Our systems were struggling to cope with increasing demands. Working with key partner Sun Microsystems, which offered both CMT and high-end SMP systems (in additional to x86) infrastructures, we now have in place a heterogeneous mix of eight high-end SMP machines, combined with a compute cluster using 78 CMT-based systems with two 1.2GHz chips in each server.

The high-end cluster has over 4,000 compute threads, making it ideal for high-throughput workloads, such as computational fluid dynamics (CFD) which can consume up to 700 GB of RAM. All of the servers run the Solaris 10 Operating System, which is an ideal platform for developing and running scalable applications that can take advantage of multiple threads. What we have is a dynamic and robust technology platform that allows researchers to manipulate different computing strengths to suit each individual application requirement. Although some might argue that a large majority of the market is still choosing to use x86, we’ve found that using a multithreaded cluster has strong gains. That being said, a willingness from programmers to experiment with code application, design and implementation is needed to take full advantage of the complex thread structure within the architectures.

What the future of HPC looks like
Looking ahead, the HPC landscape will undoubtedly continue to change. The rising proliferation of data will force organizations to move storage needs higher up the IT agenda while emerging technologies, such as multi-threading and cloud computing will change the way information is accessed and licensed. Both factors will shape the HPC server landscape over time. For now, what rings true is that the future will bring change — and in order to best prepare for the unknown, regardless of vendor choice, organizations must work with the right partners to facilitate a flexible, scalable IT infrastructure that can rapidly and easily adapt to growth patterns and changes we may not be able to forecast.

Ken Edgecombe is Executive Director of the High Performance Computing Virtual Laboratory. He may be reached at editor@ScientificComputing.com.