As more companies feel the burdens of growing data demands in terms of volume and complexity—not to mention the need to derive results on such data quickly and efficiently—the chasm between what was once considered mainstream enterprise computing and “traditional” high performance computing is is narrowing.
As we’ve addressed in other parts of this special series on lessons that HPC can carry into a growing array of enterprise application areas, including those that have a range of defined “big data” problems, this merging of HPC and commercial computing has been underway with increasing veracity over the last few years in particular—directly in line with momentum around the many data movement, ingestion and processing, memory, efficiency and other challenges enterprise users face.
While HPC has always had a foothold in key commercial segments (financial services, oil and gas, government, etc.) the technologies that were once reserved for these large-scale commercial areas are filtering down to a wider base of existing enterprise entities. It’s not uncommon lately (in the wake of the hubbub around big data) to hear about insurance companies, web retailers, content and media companies and others taking notice of HPC technologies in new ways. Bill Mannel, General Manager of Compute Servers at SGI echoed this following a conversation about this HPC to enterprise leap, noting, “Key lessons that commercial and enterprise datacenters can take away from HPC is that infrastructure matters based upon your application, your data, and the quality of service expectations of customers.”
While many won’t disagree with that point, for those with complex applications, infrastructure has to matter in different ways than it used to. As Cray’s VP of Storage and Data Management, Barry Bolding told us, one of the most important lessons for the commercial segments is productive scalability. “The commercial/enterprise space understands productive virtualization, which is a type of scaling that improves utilization of resources. The area of productive scaling that HPC brings to the table is efficient, productive scalability for complex systems. Scaling to fit an HPC solution in the coming years will require efficient parallel computing (both HW and SW), efficient parallel storage (to ensure no data access bottlenecks) and scalable analytics.
Bolding says the enterprise is seeing more and more applications needs that fit this model of parallel compute, storage and analytics. The energy sector is using new, complex algorithms to do oil and gas exploration and productive scalability is key to meeting their needs. In this example parallel, scalable storage and compute are the core to solving the problems efficiently.
Another key lesson that HPC can bring to bear is adaptive technologies, he says, noting that “for maximum efficiency and TCO it is critical to match the application need to the appropriate underlying technology. This is contrary to the cloud model where little effort is made to match the underlying technology to the application.”
When asked about the infrastructure leap from HPC to enterprise, Paul Dlugosch of Micron explained that “It is the HPC industry that first meets the most critical and difficult problems encountered in scientific and technical computing and it is true that innovations in the HPC industry often trickle down into mainstream use in commercial/enterprise datacenters.” In some cases, he says, the innovations can migrate all the way down to the client or consumer space.” In short, although the HPC industry operates at the top of this hierarchy of compute capability, there are “lessons learned in the HPC industry that have practical application throughout the entire spectrum of compute capability.”
While performance remains an important metric, Dlugosch says a myopic focus on performance can lead towards the top of a pyramid where the performance crown may be acquired but the overall market for the technology developed might becomes proportionately smaller. “When performance is the only objective, important opportunities may be missed. A good example would be the disruption imparted on high performance microprocessor vendors by the emerging need for lower power processors where less compute performance was an acceptable trade off. The lesson here, of course, is that focus on high performance may miss very important innovations that are not based on processing performance.”
Performance does indeed drive all aspects of the computing industry, but a sole focus on compute performance can leave a business vulnerable, argues Dlugosch. While the HPC industry can better afford a concentrated focus on compute performance, this does not extend to other segments of the computing industry where performance is only one of several metrics that will determine overall success.
One other area where HPC and enterprise users can connect is in the realm of risk adversity, says Dlugosch. As he explained in a detailed interview:
The old adage that ‘nobody ever got fired for buying IBM’ reflects this point quite well. Of course, IBM in this case is a proxy for any well established, mature and stable technology provider. While it may be true that nobody gets fired for buying tried and true technology, entire businesses can fail because they did not recognize important technology inflection points that were coming their way. There are many popular examples that include Wang Computer (client based word processing), Digital Equipment (personal computer) among others.
The HPC industry is quite used to operating in the domain where the opportunity for failure is high. It is the nature of pushing the boundaries of computing capability. So what lesson might the commercial/enterprise data centers learn for the HPC community in this respect? You must be willing to explore technologies outside the comfort zone defined by incremental or evolutionary improvements. Customers have a long history of driving suppliers and service providers along predictable paths of incremental improvements.
While this may be safe and meet the needs of the immediate business, following this safe path may lead to a missed opportunities afforded by new and emerging technologies. In particular, low end disruptions enabled by new technologies can be detrimental to businesses that are caught off guard. While the HPC industry is naturally focused on the high end of the computing spectrum and have a higher tolerance for risk, commercial/enterprise data centers must also take ownership for innovation and not assume it will come from their technology providers or through customer demands.
The problem of choosing the proper system for a given workload is not just an HPC issue. However, according to some, including Bill Dunmire, Senior Director of Product Marketing at SGI, “High performance computing is generally unchartered territory within enterprise data centers. It is here that “clusters” are utilized for HA (server failover) or server virtualization (e.g. V-motion) as opposed to parallel computing. Shared-memory systems are completely unknown.” He notes that in such cases, “IT will be required to develop expertise in HPC and will need to avoid inefficiencies in performance, scalability, and cost as LOB demands grow.”
Add to that general view, the more complex matters of system design and architecture which, as Jack Dongarra of Oak Ridge National Lab and the University of Tennessee told us, leads traditional HPC and enterprise users of advanced computing to two key questions—first, how can/should the internal architecture of HPC systems be changed to make them more suitable for data driven commercial applications? Second, how can/should external storage systems and their interfaces be adapted in order to efficiently orchestrate, as part of the overall workflow, the movement of data into and out of these systems? At this point in time, however, these questions seem to only generate more questions rather than any widely accepted (or even plausible) answers.
“Issues of interoperability are closely related with fundamental questions about the architecture and codesign of hardware and software infrastructure,” Dongarra explained. “Unfortunately, these same factors tend to make them relatively intractable. For interoperability has to mean more than just “everyone adopts the same standard or the same interface.” Aside from cases where de facto or de jure monopoly power is exercised, a viable approach to interoperability for infrastructure means designing protocols and interfaces that people voluntarily adopt because they can use them to achieve their functional goals while also achieving deployment scalability and sustainability over time.”
Echoing Jack Dongarra’s questions and potential roadblocks to widespread changes in enterprise computing, HPC researcher, Dr. Kirk Cameron of Virginia Tech explained that “The problems of scalability, speed, and complexity manifest acutely at the extreme scales that challenge the HPC community daily. Thus, the incessant need in HPC to maintain competitiveness by pushing simulation fidelity and scale to solve problems of grand importance to a myriad of sciences ensures the rapid adoption of cutting edge technologies.” He points to certain technologies, such as the Cell Broadband Engine, are vetted and then only briefly embraced by commercial enterprises. Other technologies, such as general purpose graphics processing units (GPGPUs), are vetted and ultimately adapted and integrated into the mainstream as evidenced by Intel and AMD embracing systems-on-chip technologies with GPGPUs built in. Much like high-performance car racing drives advances in automobile efficiency, HPC pushes the limits of computing so that commerical/enterprise datacenters can adopt best-in-class techniques and technologies to reduce the burden on their in-house R&D efforts.”
The central question is which technologies will enterprises seek and adopt that filter from HPC, especially with some of the potential barriers Dongarra and others have mentioned. To arrive at a more thorough answer to that question, we’ll be exploring a few aspects of these topics in coming special sections in the HPC to enterprise series around accelerators, HPC clouds and overall workflow/software issues later this week.