Are Data Centers Still Relevant?
With the tremendous growth and adoption of cloud services over the past decade, are data centers still relevant, especially in higher education? The answer isn’t simple, and one that institutions will have to determine for themselves. Let’s face it, data centers are expensive to run and manage. Life cycles for power, generators, and cooling in particular can be challenging on flat IT budgets. The lure of the cloud is paying for what you use, it allows smaller institutions to compete with larger institutions by not having to invest large capital dollars on building or maintaining data centers.
In recent conversations with colleagues and experts in the industry, it was noted that beyond the usual suspects such as Amazon, Google, Microsoft or other cloud providers, there hasn’t been much growth in building new data centers, except in one area, High Performance Computing (HPC).
High Performance Computing Data Center Challenges
HPC is driving the need for new data centers that can handle the cooling and power generated from GPUs used for research computing. Institutions looking to expand their HPC deployments are finding limitations in their current data center architecture and design. Traditional cooling method costs are high, and, in many cases, there is limited physical spaces along the walls to install new Computer Room Air Conditioners (CRAC) or space on the roof to place the dry coolers. Often, research grants expect existing infrastructure in place to support funding of compute nodes. For large grants, augmenting the existing data with cooling and power to support large quantity of GPU nodes can be costly and time-consuming results in grants not being pursued or obtained.
HPC Data Centers of the future
Data centers are still very relevant in Higher Education, especially when it comes to HPC and research computing. Ideally, you’ll want to look to invest in future proofing your current data center with sustainable cooling and power solutions. In terms of data center cooling, decide what are the ideal temperature and humidity ranges that your systems and equipment can handle. Look for cooling architectures that support that range, whether it rack cabinet level cooling, free air cooling, or mixture. ASHRAE has defined standards for (recommended and allowable levels)[https://www.datacenterfrontier.com/special-reports/article/11431300/understanding-data-center-temperature-guidelines]. Typically, data centers have used AC power architectures but with HPC dense racks that range from 30-100+ kW, you’ll want to investigate DC solutions as they gain in popularity and become supported at the system level.
If you are in a position to build a new data center, look for a modular design that allows buildout as you grow versus building the entire data center at once. Consider the requirements for redundancy and tiering level that needs to be achieved. If no university business services will be conducted out of the data center, build redundancy into the infrastructure, not at the system layer.
When expanding existing or building new, never take on this type of project without engaging with engineering consultants. Start with a feasibility study based on your requirements, needs for HPC, tiering level you want to achieve, and power usage effectiveness. While in the end, you may settle for some traditional models for data center, take this opportunity to be innovative.
Cloud Computing for HPC
By no means am I advocating for no cloud services, in fact, I’m a huge proponent for using cloud services when it makes sense. I’d say more of a cloud balanced model. Absolutely cloud can, should, and has been used for HPC.
Cases such as when researcher need large amounts of compute for short time spans, cloud computing works extremely well from a performance and cost perspective. In these cases, purchasing, installing, and running equipment in the data center isn’t effective. However, running intense compute 24×7 over a long period of time can be costly in the cloud. In those circumstances, on-premises provides the most value.
Security is another barrier that often comes up regarding HPC in the cloud, specifically around classified or HIPAA data. While over the recent years, all cloud services providers have increased their security offerings, identify HIPAA compliant services, and will work with IT departments to secure their data. Often, it’s just easier to keep it on-premises because of the perceived notion that it will be more secure there. That’s not always the case as intuitions still need the same controls in place and there is always the human factor. Lastly, there are still some data sets that just are not approved for cloud.
Faculty comfort level with using cloud services for HPC also must be considered. Running workloads or experiments on-premises that don’t produce the correct results or fail, often times don’t cost much more than their time. With cloud pay for what you use model, those experiments cost real money, which results in the researcher really being sure that all their code and data is correct before pushing that button.
Finding the right balance
When it comes to relevancy of data centers in higher education, HPC seems to be driving force behind the needs of maintaining, expanding, or building new data centers. Seek opportunities to build more sustainable solutions for cooling and power. Understand your research compute and storage needs as some can be handled by cloud computing. Find the right balance for your institution.