Issue link: http://cp.revolio.com/i/833286
20 | THE DATA CENTER JOURNAL www.datacenterjournal.com I n reality, bare-metal dedicated servers and colocation are an essential part of the enterprise infrastructure ecosystem. No one would deny that cloud plat- forms offer advantages: being able to spin up servers on demand is useful. In some scenarios, cloud elasticity is desirable. But experienced CIOs are well aware that in many cases, coloca- tion is the better choice. e cloud is great for startups and companies without the capital to invest in the infrastructure they need to support their operations. For established companies, however, the math is a little different. ey can af- ford the capital investment, and they oen choose to take advantage of the superior performance and reliability of bare metal. Time and again we've seen companies make the journey from the cloud to bare metal. As they grow, they develop the expertise to manage servers and infrastructure, realizing that cloud oen doesn't make sense financially. It's not just large enter- prises that make this journey, though. Smaller companies that don't want to pay through the nose for guaranteed performance and support on cloud platforms find that deploying a couple of dedicated servers in a world-class data center slashes costs. Doing so also allows them to offer a faster and more reliable experience to users. My purpose isn't to bash the cloud or cloud vendors. e cloud has a place in the infrastructure portfolio of companies of all sizes. But I've too oen heard from CTOs and business owners who were misled into thinking that IaaS or PaaS were the solutions to all their problems, only to find that the costs, performance and complexity weren't what they'd hoped for. moderN applicatioNS demaNd optimal performaNce For applications where raw processing power and low-latency IO are of primary importance, bare metal will always be the superior choice. Cloud-platform performance will continue to improve as cloud vendors iterate and bring new products to market, but in the foreseeable future, a well-designed bare-metal server will outcompete the cloud dollar for dollar. Big data and machine learning are supposedly perfect applications of the cloud, exceeding bare-metal serv- ers. On the surface, that claim makes sense: machine-learning applications handle variable workloads, oen distilling huge amounts of data. Cloud elasticity is an advantage. But elasticity isn't everything. Performance, par- ticularly I/O performance, is crucial for the fast analysis of large amounts of data. Big data applications such as Ha- doop and Apache Spark can consume all the I/O and processor performance users can throw at them. And, as com- panies that have tried to use the cloud for big data analysis have discovered, public-cloud platforms don't offer the low-latency I/O performance needed to churn through vast amounts of data quickly. Some cloud platforms are specialized for workloads of this type, but using them at scale is oen more expensive than deploying equivalent bare-metal servers in a colocation data center. If a project depends on low-latency I/O or guaranteed and predictable processor access, colo- cated bare metal is almost certainly the best option Not to mention applications that benefit from parallel processing on GPUs: the cloud is catching up in that area, but building and colocating hardware optimized for specific busi- ness needs is almost always a better fit. outSourciNg iS tHe rigHt move It's worth thinking about the primary motivation for adopting the cloud and colocation. In both cases, companies seek to outsource data center management to a third party. In the modern infrastructure world, that approach makes perfect sense. For all but the largest companies— Google and Facebook being obvious examples—the economics of building a world-class data center are unvi- able. Multiple redundant network connections, redundant power with backups, and cutting-edge security: these features are par for the course in modern colocation data centers, but they're enormously expensive for any company without deep pockets. At the data center level, it makes economic sense to share the expense among multiple users rather than have each company build its own data Given the cloud marketing frenzy of the last few years—particularly for infrastructure-as-a-service (IaaS)—you'd be forgiven for thinking that in 2017, the cloud is a foregone conclusion for large-scale infrastructure.