From rendering special effects to transcoding, Flash based storage can be an offer a revolutionary step for many use cases. Jason Danielson, Media and Entertainment Solutions, Product and Solution Marketing at NetApp provide us with a short history and roadmap of the use of flash in storage arrays in the media industry.
The current storage array landscape consists of three primary configurations:
1) All-flash arrays – solid state disk (SSD) only
2) Hybrid arrays – combination of SSD and hard disk drive (HDD)
3) Traditional HDD arrays – HDD only, perhaps with flash in the controller
Today there is a place for all three. Each serves a given cost/performance or cost/capacity requirement. On the extreme ends of the bell curve, there are the media workloads and business requirements that gravitate toward SSD- or HDD-only configurations. Highly random, latency-sensitive workloads benefit from SSD, while sequential workloads such as high-bit-rate production or those requiring large capacity benefit from HDD. Data centers usually have the need to serve both workloads. In such a scenario, which covers the majority of uses, a hybrid solution prevails.
It’s worth looking back at how we got here to understand where we might be going. While SSDs, with a capacity density to rival HDDs, have only been around in the last decade, the idea of non-volatile random access memory (NVRAM) goes back to the 1960s. And the use of NVRAM to accelerate HDD arrays goes back to the early 1990s.
Flash in the Controller
Twenty-some years ago, when disk drives were really slow and memory was really expensive, a startup storage appliance vendor, NetApp, set out to build a storage appliance that would leverage the low latency of memory to improve the performance of low-cost HDD arrays. NetApp integrated journal entry into the controller. This innovation enabled incoming writes to go directly to dynamic random access memory (DRAM) and logged to NVRAM, and then the write acknowledged back to the source. In other words, data was written to the flash within the storage system and available to be read all before the slowest part of the storage system – the disks – got involved in the equation.
Of course eventually the data needed to be written to the much slower disk drives. To speed the disk write operations, the NetApp file system coalesced the writes in memory and flushed them to disk in large sequential stripes across all of the available disk drives, maximizing the performance capabilities of the HDD array and limiting disk contention.
Removing mechanical hard disks from the acknowledged write path, thereby essentially enabling writes to memory, dramatically reduced write latencies. While this system was not a hybrid storage array by today’s standards, as the memory used was not flash (flash was too expensive and not dense enough at the time), it certainly was a hybrid storage array because it used both memory and HDDs for their unique benefits. Over the ensuing 20 years NetApp added Flash Cache – now supporting up to 16 terabytes of flash in the controller to provide similar low-latency write confirmation. More recently Flash Pools (SSDs) were introduced as an SSD storage tier with automatic tiering to HDDs to again provide low-latency response times for an even greater amount of data.
Eventually, the ubiquitous use of flash in lightweight laptops, smartphones, cameras, audio players, video games, and tablet computers drove down the price of flash – from roughly $50,000 per gigabyte in 1991 to $100 per gigabyte in 2005 to around $5 per gigabyte more recently. The last decade also saw rapid advancement in bit density, making flash comparable to HDDs. In other words, you could store a gigabyte of data in roughly the same amount of space in flash as on HDD. This factor is why flash has suddenly become so disruptive, in today’s storage infrastructure discussion. The Storage Networking Industry Association explains it this way:
Flash capacity efficiency versus DRAM
a. ~5x better $ per GB
b. ~40x better power per GB
Flash IOPS efficiency versus HDDs
c. ~40x better $ per IOPS
d. ~600x better power per IOPS
The advancement in flash-based SSDs now allows gigabytes of flash to slide into standard HDD storage enclosures, meaning that the low-latency, high-performance, energy-sipping benefits of flash can be employed in new architectures. And while SSDs are much more costly per capacity than HDDs, their use for certain workloads is financially justified.
The concept of enterprise flash drives emerged in 2008, and in 2009 Micron Technology announced an SSD with a SATA interface. What followed over the next five years was a furious run by startups to ship low-feature, high-performance all-flash arrays (AFA).
AFAs have quickly become a billion-dollar niche in the storage world. CIOs charged with updating or building data centers are looking to deploy AFAs wherever it makes technical and financial sense. In services that require millions of IOPS, such as random-access, I/O-intensive database workflows, AFAs are being deployed for all the right workloads – intense random reads, sequential reads after random writes, low-read latencies, and datasets needing to reside in memory.
The early startup vendors achieved initial success as early adopters began testing and implementing with AFAs, but for the most part they failed to create any defensible product differentiation. That’s because packaging readily available SSDs into array product offerings wasn’t enough. Further growth is not necessarily going to go to the AFA startups. Enterprise storage vendors are well-positioned to absorb the bulk of the market quickly by offering AFAs with all the enterprise features their customers have come to rely on. According to Gartner Research, Violin Memory, a startup that led the AFA market in 2012, dropped to third place in 2013 with only 20 percent year-on-year growth, while the overall AFA market tripled.
The industry is past the wow factor of increased performance. Now that data centers have adopted AFAs, and array capacities have increased into the hundreds of terabytes, IT departments are now looking to deploy AFAs with all the enterprise features found in traditional HDD arrays — such as inline deduplication and data compression, snapshots, automatic replication, easy installation and administration, etc. These features are crucial for mission-critical applications. And IT departments want to be able to control AFAs with the same tools they use to control the rest of their storage infrastructure.
Hybrid Flash/Hard Disk Drive Arrays
Successful AFA deployments in the data center don’t necessarily translate to successful media workflow storage deployments. SSDs can provide five to 10 times as much bandwidth per drive as HDDs, and for spindle-bound storage systems, that difference is a big factor. But the trade-off is not as simple as that because bandwidth per dollar — not bandwidth per drive — is usually the critical factor. While it’s possible to get the aggregate bandwidth necessary for a media workflow with AFAs, AFAs won’t improve the bandwidth per dollar, and they certainly won’t give you the capacity per dollar that you would get with an HDD system.
But there is a middle way that is very promising and that is Hybrid SSD/HDD storage arrays. There are pros and cons to flash storage much like there are with various types of HDDs.
Flash Tier Advantages
1) Fast random I/O for small blocks
2) Low read and write latency time
3) Low power consumption
4) Low noise
5) Better mechanical reliability
Flash Tier Disadvantages
1) Very high price per capacity
2) Limited capacities
3) Slow random write speeds
a) Erase of blocks
4) Slow sequential write throughput
The same period of rapid advancement in flash manufacturing also saw advancements in tiering technologies, which help to embrace the advantages and address the disadvantages of each type of storage. Combining a flash tier with a larger-capacity HDD tier in a hybrid array allows intelligent use of SSDs that can lessen the overall cost of the arrays while improving the price per bandwidth and the price per capacity of the overall system. The performance of the overall system gets a boost because hot data sits in nonvolatile random access memory while the bulk of the capacity sits in lower-cost, slower HDDs. Storage systems previously made entirely of 10,000-RPM SAS drives for performance can instead be designed with enough flash to provide performance across the given workload, and an HDD tier behind it made of less expensive, greater capacity 7,200-RPM NL-SAS drives.
The AFA market will continue to see demand for two types of products; low-feature, low-cost arrays and enterprise-featured products. However, increasingly the low-featured arrays will be delegated to small-appliance, single-use-case environments. As media operations continue to grow, their needs will meld with those of the IT data center, and enterprise features will be as crucial to the media engineer as they are to the CIO today.
Over the next five years, solid state technologies will have a profound impact on enterprise storage, but it’s not just about replacing mechanical media with solid state media. The architectural balance of memory, cache, and persistent storage will shift; applications will be rewritten to take greater advantage of the read/write profile of flash; and caching and tiering software technologies will evolve to embrace the benefits of flash for longer sequential reads as are needed for video editing and uncompressed playback. As these technologies progress, the industry would be wise to pay attention to the details and not to the hype.