• There's a principle called data gravity which says that data tends to collect and accumulate in the same places, and value tends to accumulate around the places where that data is. It's much easier and cheaper to move the code close to the data than it is to move the data close to the code. This was made apparent even on high-speed networks when Hadoop began to take off five or six years ago. For companies trying to build solutions around many terabytes or petabytes of data in a single project, it's much easier to keep things in the local datacenter if that's where the data is. Many scientific computing environments like CERN or LLNL can produce over 1 TB of data every second -- you can see Lustre's case studies around these if you want more details -- and unless there's some major fundamental innovation in the way the Internet works, it will never be feasible to move this data into the cloud.
  
• With cloud, you lose a lot of control and a lot of visibility. This is untenable for many organizations for philosophical reasons. In practice, for common application architectures, it really doesn't matter.
  
• As above, cloud application architectures are expensive to develop. Most ISVs haven't moved towards a cloud model yet; those that have tend to be monetizing their product with SaaS anyway. These applications benefit far more from more reliable hardware than the cheap hardware at scale used in the cloud.
  
• Many environments have hardware requirements that preclude the use of cloud as it exists today. Real-time systems, especially in domains like high-frequency trading, are likely to stay as close to the metal as possible, and geographically close to the exchanges. Other systems perform an order of magnitude better using boutique or specialty hardware that is not available in the cloud, like the Intel MIC (Xeon Phi) or specialty GPU hardware.
  
• At the end of the day, cloud providers are companies trying to make money. They're probably better at realizing economies of scale than you are, but when you cut out the middleman, maybe not, especially in areas like education that are very good at getting very deep vendor discounts. Cloud is cheapest when you need periodic scale and you don't need to support 100% of the load 100% of the time. As you move closer to 100% utilization, these economics evaporate a little bit. This makes cloud often a very bad choice for things like long-running, high-usage HPC clusters when you have the staff on hand to support it.
  
• If you've ever read The Phoenix Project, you're very familiar with the idea that project staff and operational staff often overlap, and when they do, this poses a tremendous schedule risk to the business. One of the biggest benefits of cloud is that it removes a huge amount of operational uncertainty from the business. If the company is large enough to have separate infrastructure and application teams, they're able to realize a lot of these benefits without paying a middleman.
  

Vulture Culture fucked around with this message at 15:35 on Apr 27, 2015

Dr. Arbitrary fucked around with this message at 17:57 on Apr 27, 2015