Multi-million user social networks, cloud hosting, Internet search and Big Data problems such as meteorology, complex physics and business informatics, all share one basic need - they each require incredibly large, complex and varied computer platforms. However, a common requirement across these systems is to "optimize the unit cost of computing." At this degree of hyperscale computing, the network, system, software, facility, and maintenance all add up to 10s or 100s of millions of dollars per project, and optimizations of a single element or the coordination of multiple elements can save the business millions. A good example of this holistic approach is Facebook's OpenCompute project, which saved the company 38% in efficiency and costs 24% less in build expense.

Similar to the automobile industry, where the racing technology from Indy, F1, and NASCAR end up in passenger vehicles, the hyperscale compute innovations we're seeing in juggernauts like Facebook will end up as line-item part numbers from vendors that are available to everyone. The timing couldn't be better, as solid state hard drives are becoming affordable and most enterprises are ramping up private cloud initiatives within their firms.

In a hyperscale design, premium computing constructs (like those seen in blade systems) are normally abandoned, favoring stripped down commodity designs that do the job at a fraction of the price. Because of the size of the deployment, rewriting an application to take advantage of the commodity compute fabric, or moving a task that was done in purpose-built hardware into custom software (e.g., disaster fault recovery), becomes cost-effective. Essentially, the decreased investment in hardware funds the software investments with ease. So what design elements are being abandoned in favor of hyperscale computing?

An example of the complex monolithic system that is being abandoned

• Premium storage array networks with expensive optical connectivity and recovery features are being replaced with a mix of locally attached and network-attached storage, eliminating the heavy burden on the storage network
• Dedicated compute, manage and storage networks are being replaced, favoring virtual LANs that reduce cabling and network costs
• High cost per port network switching is being replaced, favoring commodity network components
• High cost per socket blade systems are being replaced, favoring commodity compute components
• Devices for monitoring and management are being replaced, favoring software tools and thoughtfully architected applications
• Hot-swappable devices for high availability are being replaced, favoring streamlined hardware configuration
• Redundant power supplies are being eliminated

The best visualization for this kind of unit cost of computing design is the Google Platform from 1998 that integrated individual parts without the purchase of machine cases.

Previously, creating the best optimized hyperscale compute fabric meant that a full staff of hardware/network/applications/systems/facilities engineers was needed to drive out the costs. Today, there are firms that are using hyperscale designs to create private cloud solutions affordable for small to medium-sized business markets or for business units in large firms. Companies working in this space aim to create the highest performance per IOP private cloud solution, delivering highly scalable infrastructure solutions.

Ideally, this architecture comes in the form of a single unit that uses converged networking, a mix of local and network-attached storage, and management software included in a small form factor. There are a handful of innovative vendors offering these solutions today. Customers adopting this type of solution enjoy an extremely low-cost commitment as a minimally configured system is capable of running a base level of virtual machines in a private and dedicated system with the potential to scale as needed. Hyperscale designs also work well in large-scale deployments, where 100,000s of virtual machines are being run.

The mid-scale cloud market, comprised of 10,000s of virtual machines, is also an interesting space. Currently, mid-market integrated private cloud offerings require large upfront costs and ongoing operational costs for dedicated staff to manage and maintain the complicated compute, storage, and networking, in addition to the expensive per socket and port hardware. Buyers in this space should certainly be asking vendors the cost per VM and the cost per terabyte of storage before they purchase, as well as determining the skills that are required to maintain an infrastructure of that kind. At this point, the mid-scale solutions look obsolete, as evolving hyperscale formats require lower cost commitments, and deliver high price performance coupled with compute, network and storage cooperation.

When discussing application considerations, hyperscale architecture is a natural platform for applications designed to leverage its key features - horizontal scalability (for high throughput and increased performance), and redundancy (for high availability and fault tolerance). Earlier hyperscale architectures, as mentioned earlier, took a different approach toward performance and reliability. Data access performance and high availability relied on premium storage array networks with expensive optical connectivity and recovery features. Compute performance relied on a high cost per socket blade system and high cost per port network switches.

The service orientation and "assumed failure" approach to cloud applications puts the burden of performance and reliability assurance on the application architecture. By constructing applications as a collection of loosely coupled services, greater performance can be achieved by distributing and replicating services horizontally across commodity compute, network, and storage components. High availability can also be achieved in a similar fashion by replicating application services across the hyperscale environment and introducing a failover mechanism to mirrored services upon service failure detection.

It's important to note several additional benefits achieved by this synergy between the hyperscale architecture and applications designed to leverage it. From a performance standpoint, system monitoring software can easily be configured to detect business policy-driven performance thresholds and automatically scale or contract services based on such policies. A similar strategy can be established for high availability policies. Should the number of redundant backup services fall below a certain threshold, additional backup services can be launched before any danger of service disruption is reached. Without going into exhaustive detail, it's clear that another hyperscale benefit is the ease in which applications and platform components can be patched and replaced without service disruption. Finally, the same mechanism by which patches are applied and platforms are replaced makes it easy to test and launch new features in line with the company's business strategy.

In conclusion, organizations across a wide variety of markets require robust servers with high density performance at an affordable entry price for all levels of businesses. A hyperscale architecture combined with well-designed applications provides enterprises with a powerful tool to operate an agile business, staying ahead of the competition and exploiting new business opportunities to its advantage.