Few today would doubt that flash memory is far more efficient than trying to scale performance with the humble disk drive, but the cost of flash has caused many companies to reserve it for enterprise servers hosting their most critical applications and databases.
With flash memory devices available in 2013 from $3.89 per gigabyte and lower, hyperscale companies can now consider using flash to power entire hyperscale data centres in addition to mission-critical applications.
Flash for Hyperscale
With reliable flash memory performance now affordable enough to scale out across an entire data centre, hyperscale companies can finally take advantage of application acceleration capabilities that were initially reserved for enterprise customers. Some of these web and cloud companies, like Amazon, Google, Facebook, Yahoo and Twitter, not to mention similar companies operating in other regions of the globe, run the largest data centres in the world and rely upon massive amounts of hardware to keep petabytes of data moving at the speed of click.
Given the massive amount of data these companies have in flight, hyperscale environments are often custom-built with open source software running on bare bones hardware that is refreshed every 2-3 years, rather than the 4-5 year refresh cycles seen in enterprise data centres. Where enterprise environments accelerate applications by caching on high performance server-side flash memory, or virtualise to get full use of their servers’ capabilities, hyperscale companies typically avoid expensive SAN systems, implementing architectures that allow them to meet performance and availability needs by distributing data across vast numbers of servers.
With flash memory platforms now architected to meet the distinct needs of hyperscale environments, even emerging cloud and web leaders can now affordably deliver peak performance to their customers, many of whom become frustrated when forced to wait more than two seconds for a video to play or a page to load. These flash powered performance improvements mean companies can meet user expectations, even when content suddenly goes viral.
Getting More “Miles-Per-Gallon” in the Data centre
In addition to performance, hyperscale companies can now take advantage of the ecological benefits offered by flash memory. This is a significant improvement as we scale our data centres to keep up with today’s data demands. As we connect more devices, these demands are skyrocketing, with Wikibon noting that in fact, we will soon be creating five billion gigabytes of data every ten minutes. Many communities are not embracing the idea of having new data centres built in their backyards, creating the need for new solution to deliver more performance in a smaller footprint.
The newest flash memory platforms for hyperscale use less energy to power than disk, around 25 watts, while providing terabytes of storage. Cooling costs can also be reduced, often by as much as 50 percent compared to scaled-out disk infrastructure. When you calculate the performance achieved per watt and the terabytes of flash memory density now available, it quickly becomes apparent that flash delivers more byte for your buck compared to disk.
Flash as a Memory Tier Through Open Source APIs
All these flash powered benefits now available for hyperscale can be further extended by unlocking flash’s full potential with new APIs. Through new software capabilities, flash can now be used as a new memory tier that can offset the need for massive amounts of DRAM. Traditionally, DRAM uses much more energy and has a cost 10 times that of flash, gigabyte for gigabyte. DRAM also uses much more energy to operate than flash.
Even SSDs fall short delivering on this potential, as they often utilise a “backwards compatible” architecture with the disk drive, and introduce more latency and complexity by requiring multiple controllers to manage higher capacity SSDs. New media like flash demands new programming primitives to ensure maximum efficiency, and without a seek head on our digital flash platforms, there is no need to keep disk era code in our 21st Century data centre architectures.
With these new APIs for flash, companies can utilise the flash memory tier to advance hyperscale computing. While the APIs need to be inserted into software code, given that hyperscale companies frequently embrace open source software, these new APIs are expected to first emerge in applications used for hyperscale computing. Already, developers are finding that these programming primitives are helping them get updates to market faster while delivering on increased performance and enabling new features.
The Displacement of Disk
Despite these exciting innovations, it’s not time to ditch all disk storage just yet. Similar to how tape stuck around for decades after disk made its first appearance (not to mention its recent resurgence for archival storage), disk will still make an ideal medium for backups. However, its place as primary storage in enterprise and hyperscale environments is diminishing as flash continues to become more affordable for all the needs in the data centre, and as the industry adopts the new protocols possible when treating flash as a memory medium.
It will be exciting to see what hyperscale companies - industry leaders as well as tomorrow’s trailblazers - are able to achieve with all-flash data centres. Whether streaming entertainment online, developing new social media tools, powering e-commerce, or introducing an entirely new data-driven innovation, 2013 will be a year shaped by breakthroughs in our connected experiences powered by flash on the data centre side of the Internet.
Posted by Ajay Nilaver, Vice President of Products at Fusion-io
Ajay oversees the development of flash memory products that fuel all-flash data centres. You can follow Fusion-io on Twitter via @FusionioEurope.