Many data centers have made an initial foray into flash-based storage.
The first stage of flash adoption is sometimes motivated to solve a performance problem for a specific application. But, frequently, end users become so enthralled with the performance of flash that they begin demanding it across a broader set of applications.
In this article, you will learn the variety of flash investment options in the market and how to make the transition from purchasing a point flash product to a data center-wide flash product.
Flash precision strike
Flash storage products are available in a variety of implementation options. From server-side flash PCIe cards, DRAM memory modules and drive form factor SSDs, to network and array-based flash storage systems. Each technology has its own distinct advantages and disadvantages, but what is more important is how these products can scale across the data center. Flash is significantly more expensive than hard disk drive storage on a cost per GB basis, so IT planners must ensure their investment in flash pays off.
Many organizations dip their toes in the flash waters by deploying server-side flash products. This represents a highly targeted approach since the flash resource can be installed right where the application lives -- on the host. What's more, server-side flash offerings from companies such as Intel, HGST, Virident and Toshiba typically provide lower latencies than other forms of flash because there is no storage network for read and write IOs to traverse; instead data passes over the backplane of the server motherboard.
One of the challenges with a server-side flash investment is the flash resources are captive to the host. If the host only uses a sliver of flash capacity, there's potentially a lot of expensive storage that's going to waste. This may not be regarded as a significant issue if flash is installed on just a handful of hosts, but if there are dozens of servers each with their own flash resources, low utilization will drive up the cost of flash ownership significantly.
Another major functional consideration with regards to server-side flash is how to best support virtual machine (VM) migrations. If a VM on a host with flash installed needs to move to another server, the administrator must make sure there is a flash resource with available capacity to host that VM before the migration takes place. Otherwise, users will experience a drop in performance. Keeping track of flash resource availability in this manner is not only cumbersome and time-consuming, but it could ultimately impact business agility and application quality of service (QoS).
Array-based flash products allow data center administrators to centrally manage flash resources across any host that is attached to the array via a network or direct connection. This alleviates many of the resource management issues described above, and it helps enable non-disruptive VM migrations. But some flash array platforms are based on traditional "scale-up" storage architectures. With these systems, you have to make an upfront investment in resources that may not be used initially. This means you could be significantly overpaying for flash if you are only using a small percentage of the system when it is deployed.
Another challenge with scale-up storage architectures is that the storage limits on these systems can potentially be exceeded. This typically requires either adding another array, which could lead to creating islands of storage, or performing a forklift upgrade to a larger scale-up storage array. In either case, this can drive up the total cost of ownership (TCO) as organizations would once again have to make another large up-front investment, and in some cases, the original storage array investment may not even be fully depreciated. Moreover, the IT organization has to allocate time to either implement a secondary system or perform a data migration to a bigger box.
However, it is important to mention that many all-flash, scale-up storage systems can provide significant performance and capacity in a relatively small footprint. Pure Storage, for example, offers an entry-level all-flash array that provides 40 TB of capacity with hundreds of thousands of IOPS. This type of system could easily satisfy the application performance and capacity requirements of many environments. And furthermore, it could enable some data centers to free up valuable floor space and save on power and cooling costs by consolidating their disk-heavy legacy SAN and NAS disk platforms. Also, some scale-up, all-flash arrays provide the ability to perform data in place controller upgrades. So, the next generation of performance can be achieved without having to buy a second storage system.
To step more elegantly into a flash environment, data centers may want to start out with a small flash footprint, but have the ability to scale out to meet business demands. Scale-out flash products provide a way to granularly add resources as they are needed. Products like EMC's XtremIO product, for example, enable storage administrators to deploy a single node of flash storage and scale out to multi-PB configurations. In addition, flash capacity and processing resources, like memory and CPU, can be added granularly rather than all at the same time. This allows storage planners to deploy the right resource at the right time without overprovisioning flash, CPU or memory devices.
Some scale-out flash offerings, such as those from SolidFire, require users to start out at 4 node cluster configurations, which provide 35 TB of usable capacity and 200k IOPS, but can then scale out in smaller incremental 10 TB capacities. These systems are priced for entry-level users and the benefit of a clustered configuration is that it provides more redundancy and higher availability than a single node deployment would. These systems scale into the multi-PB range and can be retrofitted on the fly with new flash devices for improved storage efficiencies. In addition, by allowing mixed generation nodes to coexist within the scale-out architecture, data centers essentially never have to perform a forklift upgrade to get improved storage performance or better flash economy.
Caching software manufacturers like SanDisk, Intel and PernixData provide software technologies that allow businesses to manage, share and provision server-side flash resources as though they were logically on a centralized storage system. These software products essentially aggregate server-side flash resources so data that is residing on server-based flash can be replicated or mirrored across multiple flash devices across multiple hosts. It also allows flash to be shared across those hosts that don't have local flash resources. This enhances flash utilization, data availability, enables non-disruptive VM migrations, and allows data centers to scale out their flash resources on demand.
Interestingly, PernixData recently announced the capability of its FVP software product to aggregate not only server-side flash (DRAM, PCIe and SSD) but to also manage flash resources on external arrays. This gives data centers the ability to deploy flash in any format, but still retain the ability to centrally manage and provision those resources. It also gives application owners the flexibility to choose the flash option that best meets their needs.
Organizations need to take a strategic approach towards implementing flash storage in the data center. Often in the rush to solve a performance pain point, storage and application administrators may quickly choose to implement a point flash storage product. But flash adoption will typically accelerate once the performance benefits are realized. Consequently, it is important for IT planners to consider those products that enable flash to efficiently scale out across the data center without leaving management challenges and cost overruns in its wake.
How all-flash differs from commodity storage
Comparing memory channel flash with PCIe flash storage