Research Computing prepares for emerging and foreseeable needs
Research Computing, for example, established a cluster-scale computing strategy that explicitly recognizes and supports data-intensive and high-throughput computation and information-processing. Instead of delivering cluster systems designed for traditional High-Performance-Computing workloads — and leaving other workloads to fit in as best they can — Research Computing has established a computational cluster designed for traditional HPC and a distinct cluster designed for data-intensive and high-throughput computation in addition to a cluster engineered for traditional or legacy HPC.
Likewise, Research Computing has developed and established a substantial suite of storage capabilities, recognizing that the storage and management of data is as significant as processing it, and as significant as sheer computation (e.g., modeling / simulation). In just a few years, Research Computing has iteratively developed a secure research workspace designed for computationally modest research on regulated data; it is now generally available.
These adjustments, and others not mentioned, are driven by Research Computing’s effort to orient to the institution as a whole, tune institutional cyberinfrastructure to demands from the fullest possible range of disciplinary areas, and prepare for emerging and foreseeable needs. Whereas research centers and institutes tend to have disciplinary or project foci, Research Computing’s aim is to cultivate a comprehensive service to the institution.
Research Computing’s approach to data science is part and parcel of this institutionally scoped orientation. Some aspects of data science require cyberinfrastructure that is present within Research Computing’s existing cyberinfrastructure services: the Dogwood cluster for multi-node MPI and MPI+OpenMP hybrid workloads (approximately 10,000 cores), Knights Landing accelerated compute nodes, the Longleaf cluster for data-intensive workloads (approximately 5,400 cores), graphics processing unit nodes, world class high-performance storage (approximately 2PB), general purpose persistent storage (approximately 4.5PB), active archive storage (approximately 6PB), self-service private cloud scientific workstations, the secure research workstation environment, etc.
In addition, Research Computing is embracing more special purpose cyberinfrastructure capabilities targeted to the techniques and methods prevalent in the field of data science.
1. Research Computing has formal relationships with Amazon AWS, Microsoft Azure, Google Compute Platform, to facilitate the dimensions of inquiry — common in data science — that is well suited for cloud services. The cloud strategy is explicitly to embrace the major providers so that researchers may avail themselves of the best platforms for their work, especially in cases where the provider has differentiating capabilities and in cases where the resource need is too new or emergent or comparatively unique for it to be otherwise available; some workloads will be born in the cloud, conducted in the cloud, and never leave the cloud.
2. Research Computing has procured cyberinfrastructure that is state of the art for techniques and tools used in data science. For example, three NVIDIA DGX-1 appliances are being implemented; and 16 nodes each with four NVIDIA Volta GPUs (with NVLINK) have been procured to add to the Longleaf cluster. There is no better technology for “deep learning” and other data science-oriented tasks.
3. The members of the Research Computing Engagement Team are knowledgeable with respect to the software applications and analytical methods that are commonly applied in data-driven research. For example, the most recent hire has advanced degrees in statistics and economics, has done advanced study on machine learning, and is actively engaged in projects that are aptly described as data science. Further, knowledge and experience in data science are core expectations for new recruitments for Engagement Team additional or vacancy-filling members.
As data science has become a core or even essential component of nearly any research inquiry, so too has the cyberinfrastructure that supports the broad range of research inquiry at the University become a direct enabler of data science. The Research Computing cyberinfrastructure strategy is arrayed as an institutional push in the varied realms data science encompasses.
In short, Research Computing sees cyberinfrastructure and expertise salient to data science as a strategic and significant domain in which to focus and blend institutional resources, just as Research Computing blended resources from traditional HPC cluster computing into data-intensive computing and storage.