Server and network advancements have commonly outstripped storage device’s ability to keep up with them. Data centers would buy powerful new arrays and attempt to keep them performing with fast-growing data by using more spindles and high performance disk clusters. This helped, but sooner rather than later the data center replaced the now-aging array with the fastest new one. And the entire cycle begins again: fast-growing data, accelerating IOPs, low latency requirements, and virtualized environments threatened to overwhelm HDD arrays.
Flash memory in the enterprise started to make a bigger splash with flash caching and Tier 0 flash for highest IOPs at a rate HDDs and traditional storage controllers cannot match. Soon some storage makers delivered hybrid arrays to markets that were fearful of the price of all-flash and that covered vendor investment in traditional HDD arrays. At the same time, vendors worked feverishly to develop all-flash arrays with a reasonable price tag that included enterprise storage services.
As 2013 wound down, 2014 saw maturing introductions of all-flash arrays whose prices were starting to match – or even undercut -- high performance disk-based systems. Just as important, all-flash storage was being manufactured with native enterprise storage services. Not all arrays offered it, and some arrays were more mature with richer services than others. But clearly, all-flash arrays in 2015 are approaching wide-scale enterprise adoption.
So if the earlier question was “Is flash performance worth the price I’m paying for it?” then the new question is “Which all-flash storage is optimal for my high-value applications?” Thus the discussion is shifting from types of flash to business-oriented discussions on how flash can and should accelerate and improve applications and business processes.
Enterprise data centers are exponentially growing their storage footprint. At the same time, critical applications like big data analytics, database performance, and heavily virtualized networks require higher IOPs and lower latency. More data centers are adopting flash storage to help them meet heavy application needs, the biggest drivers including:
Growing acceptance of flash storage advantages.
Flash’s dramatically higher IOPs performance and much lower latency.
Lowering prices for flash, especially against the cost of high performance HDDs.
New flash-based enterprise storage services without the need for 3rd party products.
Flash storage fulfills these needs and more but flash buyers face some difficult choices. First of all, flash storage options include all-flash arrays (AFA), all-flash appliances, server-level flash, and hybrid flash. Each option can work in the right environment, but for storing Tier 1 production data in a high performance environment, all-flash arrays are the way to go.
First let’s look at the flash storage alternatives to AFAs.
Flash appliances. Flash appliances are built from raw NAND chips instead of SSDs which are usually NAND flash with flash controllers and other components built into housing. The flash appliance maker adds the controller software outside of the flash and claims very high performance. More flash appliances are offering enterprise services in order to hit the bigger Tier 1 storage market but many of those options need to be turned on separately and can be very expensive.
Server level flash. Server-side flash with PCIe cards serves ultra high-performance, niche applications. As an application-specific flash solution it is not viable for production data in the data center. In fact, an AFA with high performance, low latency and the ability to optimize multiple workloads can serve even the highest IOPs/lowest latency applications, which will save the significant cost of server-side flash.
Hybrid arrays. Most hybrid storage products are traditional HDD-based storage systems with the option to integrate solid state flash memory in a disk drive form factor. Some hybrid arrays are still in the market but with drops in all-flash cost and the addition of storage services, there is little reason to invest in hybrids. Even HDD vendors who offered hybrid arrays primarily did so as a stopgap measure while developing all-flash products.
The Difference: All-Flash Arrays
AFAs were traditionally used for high-performance storage when traditional flash caches or hybrid arrays did not provide sufficient performance, and when users wanted storage services that were native to the array. These arrays are all about optimizing flash around data dispersion and reduction, wear leveling, zero write amplification, and more. Qualities like density, performance, storage capacity and reliability/high availability push AFAs deeper into the storage infrastructure mainstream.
Here is how it works: flash controllers are the key to performance in any flash architecture. Controller chips do the heavy lifting to manage raw flash at the chip level: writing data, managing deletion and garbage collection, and handling all-important error correction.
Flash appliances create proprietary flash controllers to manage raw NAND chips, usually writing them in an FPGA or ASIC. That gives the vendor ultimate development control but also saddles them with proprietary development cycles. The vendor is solely responsible for upgrading their proprietary controllers to keep pace with flash industry advancements. When an AFA vendor uses industry-standard SSDs in their array, they can leverage industry-wide developments as well as their own internal R&D. This enables AFA vendors to continually upgrade their SSDs without a correspondingly expensive controller development.
Cost Considerations
All-flash array costs are dropping but so are HDD costs, leaving AFA the more expensive alternative. Many customers look to the capital cost price tag alone as their cost consideration. Capital expense is an important factor in the cost equation but does not exist in a vacuum.
First, some AFA product costs go far beyond the simple array purchase price to include training, deployment, configuration, and extra support costs.
Second, consider application productivity gains from increased performance in the cost savings category.
Third, operating expenses should be far less with the AFA’s smaller footprint and energy efficiency.
Fourth, consider management savings. If the AFA has a rich feature set and is simple to manage, then even generalist IT will be able to successfully manage storage with a minimal investment of time and expense.
The combined total of the above is the true total cost of ownership (TCO): the cost of the entire customer experience throughout the AFA’s lifecycle.
Buyers will still want to carefully weigh costs between competing vendors. For CapEx, look for a system that will meet your needs for the next 3-5 years. Look for all-inclusive pricing including the cost of the system with storage services and required training. Treated separately, the cost of the add-ons can be more than the system’s purchase cost. Beware of low come-on prices that actually require lots of extra services and components.
For OpEx, you will get a significant break on energy costs. Several vendors proactively replace aging controllers and components as a competitive offering. However, carry out due diligence for high ongoing maintenance costs. Storage vendors across the board often offer a low support contract the first year, only to raise support costs the second year. Look for reasonable support costs that include 24x7 technical support with actual support engineers, as opposed to low-level script readers in a call-in center.
All-Flash Array Vendors
AFA vendors range from established big storage vendors to hard-charging start-ups. No matter who the vendor is, subject all AFAs to a rigorous set of feature questions including high IOPs/low latency, high availability, dedupe and compression, scalability, simplicity, the ability to handle mixed workloads, reliability, native data protection and encryption, and a high degree of automation.
High IOPs / low latency
The ability to provide high bandwidth, high IOPs and low latency is fundamental to flash performance. High performance is crucial for mission-critical applications and also accelerates business-critical applications like Exchange. This data center-wide benefit goes a long way towards adding value to the AFA purchase.
IOPs and latency are related but not synonymous: input/output per second is the total number of transactions that storage is capable of processing every second. IOPs does not specify the amount of data processed per transaction, only that there was a transaction. Latency is the amount of time it takes storage to deliver a single storage transaction, such as launching reads and writes.
Decreasing latency is particularly important because of cumulative wait time. Latency lag measured in fractions of seconds might not seem to have a big impact. But multiply that single lag across thousands to millions of IO processes per month per application. The cumulative effects can badly impact application performance. AFAs with sub-millisecond latency will significantly speed up application performance.
The best AFAs will provide high performance and low latency to a variety of workloads. Here are the questions to ask vendors about your application needs:
· How sensitive are your applications to IOPs and latency? Many business applications will benefit from high IOPs/low latency. Others require it. Other applications like voice over IP, streaming, or massive online environments need high IOPs/low latency. Large virtualized environments are also prime targets for low latency.
· What does your access pattern look like? For example, VDI is write-centric while databases are read-centric. Very fast flash reads will greatly benefit read-centric databases applications with dramatically lower latency. Write-centric applications like VDI will also benefit from flash’s far superior write performance over disk.
· What is your typical block size? 4KB is the standard block size for HDD and flash storage. Some newer AFA entrants replace 4KB larger variable block sizes up to 64KB. This enables efficient and flexible block alignment with applications.
Scalability
Look for an AFA that scales economically and non-disruptively. Non-disruptive features should include capacity expansion, controller upgrades, and software updates to maintain data without performance loss. Controller upgrades should be non-disruptive and not force a new array purchase. Industry-standard SSD form factors help to make AFAs highly scalable with features like stateless hot-swap controllers, separate expandable storage shelves, and hot-swap drive carriers that do not cause downtime when serviced or replaced.
Mixed workloads
Modern data centers have a variety of workload types generated from hypervisors, databases, Exchange, test/dev VDIs, and more. AFAs should provide high performance to all production data regardless of workload type. For example, a video file is sequential: no matter how large it is, IO is predictable and easier to read and write. Random access IO cannot be reliably predicted. Virtualized environments are infamous for random IO as multiple VMs stream interleaved IO to the shared storage array. Any type of storage will handle sequential relatively well; random access is where the rubber hits the road and where flash’s high performance has a distinct advantage over HDDs.
Deduplication and compression
Flash deduplication and compression allow flash arrays to offer high capacity. Ideally dedupe and compression should have a minimal effect on performance.
Inline dedupe and compression are critical to preserving capacity in fast-growing data environments. Here is why inline matters: reducing data as it is coming into the array greatly reduces capacity requirements. Dedupe ratios will vary depending on the application but should deliver a minimum of 4:1. Dedupe-friendly applications like databases could see upwards of 10:1.
With flash, dedupe not only preserves capacity but also improves the performance of an all-flash array. Data dispersion helps flash take advantage of higher parallelism, and reduces the writes that eventually wear down SSDs.
Simplicity
Simplified installation, training, and management relieve IT of heavy storage burdens. This is particularly important to smaller data centers that operate with a small staff of IT generalists. These data centers experience the same growth rates as the larger environments, and must manage without adding more staff. Simplicity is the best means to meet this goal without sacrificing quality.
For example, the AFA should offer highly configurable settings using a wizard and automated processes. Simplified management will include operations like growing and shrinking volumes, monitoring capacity and performance, managing hosts and host groups, and managing snapshots. Add automatic remediation for errors to dramatically shift the management burden from administrators to the storage system.
Reliability and protection
Reliability and availability depend on AFA architecture and native data protection services. For example, active/active IO handling protects against application outages; crucial to mission-critical applications. In this configuration, connectors between dual controllers transfer IO traffic from a failed controller or component to the other controller. Flash arrays that store all data writes in persistent NVRAM before writing to SSD will also ensure data consistency. RAID optimized for flash is also crucial to reliability. Far from being an outdated HDD protection feature, flash RAID actually improves performance in flash environments with much faster rebuilds thanks to uniform distribution of parity bits across multiple drives. And with flash, it is possible for a RAID design to rebuild data in a many-to-many configuration.
Data protection/DR is also vital for AFAs. Older Tier 0 flash products did not offer data protection services in the data path but things have changed for Tier 1 flash storage. New AFAs should offer native snapshots and data replication functionality. Performance impacts from snapshots are not an issue with flash, which accelerates data protection operations above HDD or hybrid-based DP. Capacity can be an issue on flash; be sure that your AFA is space efficient by deduping and compressing snapshots.
Photo courtesy of Shutterstock.
Inside TEMPlate====>
Full Site
About
RSS