Cloud computing provides companies the ability to run workloads on-demand, when they need them, and only pay for the resources used. Those resources may be related to processing power (compute) and may be infrequent or steady-state. A website will need to be available 24 hours per day. That would require a steady-state compute. If there are times when usage may spike, say at Christmas time for a retail merchant, cloud computing will allow the merchant to scale to add additional compute capacity seamlessly. The same criteria would apply for storage, database processing and any other application needs.
If cloud computing offers scalability (adding power as needed), elasticity (scale on demand), what is the difference between cloud computing and hyperscale computing? There are no good, definitive answers. Even the definition of cloud computing has a wide range of possible answers. In general, hyperscale adds global availability and, at least, the appearance of nearly infinite scalability.
What is Hyperscale?
Hyperscale computingOpens a new window generally means that a data center has massive compute power, huge amounts of storage, and extensive network fabric. In hyperscale computing, you can easily scale not just from a single server to a few but from a few hundred servers to thousands. Organizations must have compute availability and automation to provision and control the compute resources. Automation is key for hyperscale.
This level of scaling requires appropriate architecture at the data center level. Cheap, dumb servers are common and replace the more complex rack systems in traditional data centers. Many of the data centers utilize unique networking infrastructure (a good example is the Google Jupiter fabric – 1 petabit bandwidth inside a data center). And, of course, the data center must be powered for tens of thousands to millions of servers. Some data centers are rated for multiple megawatts. One of Google’s data centers in the Netherlands is contracted to use the entire supply of 62 Megawatts from a nearby wind farm.
The majority of processing and storage and most communications between processes will happen within that single data center. Redundancy will be between data centers and will involve some latency, but hyperscale computing will generally be in a single physical location. Global redundancy and availability are baked into hyperscale computing.
Learn More: Is It Time To Look Beyond Mainframes in the Hyperscale Era?
Hyperscale vs Hyperscaler
While hyperscale is a generic term, Hyperscaler usually means one of a few very large, hyperscale level corporations who own massive data centers around the world. These are usually cloud providers such as Amazon for AWS, Google for GCP, and Microsoft or Azure. But hyperscalers are not limited to just cloud providers. Facebook and Apple own massive data centers that they use to provide services and to support internal operations.
Between the big cloud providers and companies like Facebook, they maintain more than 15 million servers. These are referred to as Tier 1 or Mega Hyperscalers. Software as a Service (SaaS) operators can also be hyperscalers. SAP and Salesforce require enough resources that they are themselves hyperscalers. Estimates of resource availability among the large tier two hyperscalers are 500,000 servers across their many data centers. You can see that even in terms of hyperscalers, there is a big difference between the top tier and second tier. So, where do you fit?
Who Needs Hyperscale?
When looking at the scale of hyperscale (pardon the pun), you may wonder who, other than these huge providers would ever need that kind of scale? As data repositories grow, as businesses utilize more artificial intelligence and machine learning, as employees move more and more towards a fully remote lifestyle, the need for computing resources also grows.
Since the advent of the Internet, we moved workloads, usually simple ones, to the web. Cloud computing was the obvious evolution as processing needs grew and data became more complex. Hyperscale computing is the next evolution in the era of big data. While very large organizations like the big multinationals and national banks are already approaching hyperscale, with the advent of the Information Age, many businesses are making data their primary offering.Â
Combine the need for easy data accumulation with the heavy processing requirements of artificial intelligence and machine learning, even startups may want to take advantage of hyperscale computing.
Learn More: Stalled Cloud Migrations: Look at the People, Not the Tech
4 Steps to Becoming a Hyperscaler
The first step in becoming a hyperscaler is to invest $10-20 billion in building a new data center. Since that is a steep curve, we will take a few shortcuts.
The real first step is identifying workloads that require hyperscale. Do you ingest massive amounts of data? Transform it? AI and ML? Do you offer a SaaS and need to scale for user growth? Do you need to scale for peak processing? If you don’t need to scale out to thousands of processing nodes, you don’t need, and shouldn’t worry about, hyperscale.
Step 2 is the application and/or data architecture. Hyperscale will not be cheap. Once you have determined that normal cloud computing scales will not work for you, you will want to make sure your workloads are architected appropriately. Service-oriented applications, especially micro-services, align better to hyperscale than do monolithic applications. If your processes are based on APIs, you will more likely be successful.
If your need is primarily data-oriented, you will want to take advantage of distributed processing. If you are mainly utilizing standard relational databases or even column-oriented large-scale databases, it will not be easy to move to hyperscale. While many databases can scale quite well, thousands of nodes in a relational database will be problematic.
The third step is choosing a provider. You will most likely not build your own data center, but fortunately, there are a number of providers who will allow you to scale to hyperscale levels. Choosing a provider requires the provider to have the required resources available, but you should also validate that the tools and skills required to run a hyperscale platform exist in your team. Azure, AWS, and GCP are all great choices that offer hyperscale compute and storage. You can choose between VMs, containers, or serverless, and they all provide at least some choice of automation.
In my opinion, if you are more focused around data and want to be primarily SQL-based, GCP is the better choice. BigQuery can scale across thousands of servers transparently and is a serverless technology. It integrates well with cloud storage and makes ingestion of data and federated queries simple. For a more non-SQL oriented data focus, AWS and Azure (as well as GCP) offer file based hyperscale processing (like hadoop, spark, etc). They are all good choices with plenty of options.Â
Finally, the fourth step is automation. You cannot scale to thousands of servers/nodes without a fully automated environment. The type of automation you require is at least partially dependent on the architecture of your solution. In a serverless database environment, such as BigQuery from Google, scaling will use as many resources as you are willing to pay for. There is no need to intervene at all.
In a container environment, the application and all dependencies will be automated via startup scripting. All of the major cloud players provide multiple choices for container orchestration. VMs and file-based processing (like Hadoop) require more automation configuration than the alternatives. These work just as well, and in many use cases, even better than container and serverless configurations, but they require additional upfront analysis and, possibly, re-architecture.
Learn More: Pure Storage Extends Hyperscale Computing with DirectFlashâ„¢ Fabric
Summary
Hyperscale is an expensive and complex technology that a vast majority of organizations cannot afford or operate. It is impressive and provides massive processing power but requires massive upfront expenditure. It is much cheaper and easier to process at hyperscale velocities by utilizing an existing hyperscale provider such as AWS, GCP, or Azure. However, the latter option is also not cheap or easy. You will likely need to re-architect an application (or your entire infrastructure) to really utilize hyperscale. Lift and shift works in on-prem or the cloud, but hyperscale requires cloud-native architectures.
Is your organization ready and equipped to adopt a hyperscale computing infrastructure? Comment below or tell us on LinkedInOpens a new window , TwitterOpens a new window , or FacebookOpens a new window . We would love to hear from you.