What do Virtualization and Cloud executives think about 2011? Find out in this VMblog.com series exclusive.
Contributed Article By Peter Kirkpatrick, founder and chief architect of Aprius
Ushering in the Year of Solid-State Storage
December brings a whirlwind of nostalgia for the year gone by and a wide range of predictions for the 12 months ahead. For those of us in the virtualization market, this turn of the calendar symbolizes the end of a year of great advancement and effort. It also provides an opportunity to look ahead at the possibilities that await the market now that so many virtualization building blocks are in place. One of the most significant -- commoditized flash solid-state storage -- sets the stage for higher performance for hot data in 2011.
Flash storage evolves from the exotic to the everyday
In 2010, we saw flash storage migrate from an expensive resource to one offering usable performance at reasonable prices. This shift expands the use cases beyond high performance computing, government and military to enterprise database and file system caching. Soon, we will see flash storage used at the first tier of persistent storage. The marrying of technologies such as Converged Enhanced Ethernet (CEE), PCIe SSD flash and I/O virtualization will enable flash-based arrays and appliances to deliver high performance without sacrificing economics, scalability or usability.
This is a positive change, but one that will also affect the way in which storage performs and behaves. In response, we will need to adjust the manner by which we architect those systems in 2011. Data center systems innovators will continue to leverage the benefits seen in low latency, fast response times and an order-of-magnitude increase in I/O operations per second (IOPS) and throughput. They will further their efforts to correct the limitations related to flash storage, including cost, capacity, endurance and scalability. Finally, we will see a re-imagining of direct attached storage (DAS) and storage area network (SAN) models.
The short history of SSD
In 2010, we saw two flavors of SSD, single-level cell (SLC) and multi-level cell (MLC), but neither was completely satisfying. SLC is fast and durable, but too expensive for widespread use. MLC is cheaper, but fulfills the "you-get-what-you-pay-for" cliché. It's slower, and it just isn't a durable option. So, companies that wanted SSD were forced to choose between performance problems and major capital expenditures.
Then, several vendors introduced advances in SSD that will echo throughout data centers in the coming year. Today, MLC delivers excellent performance within enterprise-class systems at affordable price points. Furthermore, array vendors are getting smarter about how to actively manage MLC flash devices to ensure endurance and performance, playing to MLC's strengths and using smart logic and device management to overcome its weaknesses.
Raising a challenge to traditional storage protocols
In 2011, we will see ongoing challenges to traditional storage protocols, such as SATA and SAS. In their place, innovators will use novel interfaces to triple or quadruple performance outcomes. Unlike the technologies designed for spinning disk, flash storage, particularly for caching and random-access applications, will reside directly on the host system I/O bus via PCIe-style flash storage cards or the emerging PCIe SSD form factor (http://www.ssdformfactor.org/). These devices provide minimal latency and maximum bandwidth between CPU and storage, without protocol conversions. This technology is game-changing and will continue to impact the entire data center environment.
How will this affect storage systems and connected servers, in particular? While some architects will continue to force flash storage into current storage arrays, the most successful operations in 2011 will be those that disregard traditional boundaries in favor of hybrids. We're likely to see combinations including DAS and shared storage, NAS and SAN, tiering and caching.
Local access, shared storage pools
Even with all the advances in SSD in 2010, full throughput on shared flash remains limited by increased latency across multiple hops in the fabric data path, SAN protocol processing and CPU processing of both data and control path traffic. Moving past these limitations requires making flash storage available as local devices that are physically pooled and shared for scalability, efficiency, concurrent access and high availability.
Such local access with shared pools of flash storage will deliver increased application (and VM) performance, increased density, as well as the ability to virtualize I/O-intensive applications and lower overall cost per GB and IOPS. This approach further expands the value of DAS flash storage by allowing hosts to access multiple flash storage controllers simultaneously. On the back-end, we'll see many flash storage devices within converged fabrics, such as CEE, thus boosting peak IOPS for the most demanding applications, regardless of where they exist in the physical infrastructure.
In the coming year, flash SSD will become even more compelling for enterprises that seek performance and efficiency. Advances in related technologies will greatly reduce the storage I/O problems that were prevalent in 2010. SSD today is efficient, affordable and enticing, and it sets the stage for a bright 2011 in the data center.
About the Author
As founder and chief architect at Aprius, Peter Kirkpatrick has broad experience in research and product development of high speed communication systems. At Intel, within the Digital Enterprise Group, he performed path finding work in server systems architecture utilizing next generation computer bus protocols, high bandwidth system interconnects and adaptive electronic filtering. Prior to this, at Intel and Lightlogic, he also designed and produced innovative 10Gb/s system interfaces for the enterprise data communications market. Peter attended the University of Colorado in Boulder, earning a MSEE in 1999 in Electrical and Computer Engineering. He has been awarded 11 US patents and published more than 10 peer-reviewed papers.