An all-flash array (AFA), also known as a solid-state storage disk system, is an external storage array that uses only flash media for persistent storage. Flash memory is used in place of the spinning hard disk drives (HDDs) that have long been associated with networked storage systems.
Vendors that sell all-flash arrays usually allow customers to mix flash and disk drives in the same chassis, a configuration known as a hybrid array. However, those products often represent the vendor's attempt to retrofit an existing disk array by replacing some of the media with flash.Content Continues Below
All-flash array design: Retrofit or purpose-built
Other vendors sell purpose-built systems designed natively from the ground up to only support flash. These models also embed a broad range of software-defined storage features to manage data on the array.
A defining characteristic of an AFA is the inclusion of native software services that enable users to perform data management and data protection directly on the array hardware. This is different from server-side flash installed on a standard x86 server. Inserting flash storage into a server is much cheaper than buying an all-flash array, but it also requires the purchase and installation of third-party management software to supply the needed data services.
This article is part of
Leading all-flash vendors have written algorithms for array-based services for data management, including clones, compression and deduplication -- either an inline or post-process operation -- snapshots, replication, and thin provisioning.
As with its disk-based counterpart, an all-flash array provides shared storage in a storage area network (SAN) or network-attached storage (NAS) environment.
How an all-flash array differs from disk
Flash memory, which has no moving parts, is a type of nonvolatile memory that can be erased and reprogrammed in units of memory called blocks. It is a variation of erasable programmable read-only memory (EEPROM), which got its name because the memory blocks can be erased with a single action, or flash. A flash array can transfer data to and from solid-state drives (SSDs) much faster than electromechanical disk drives.
The advantage of an all-flash array, relative to disk-based storage, is full bandwidth performance and lower latency when an application makes a query to read the data. The flash memory in an AFA typically comes in the form of SSDs, which are similar in design to an integrated circuit.
Flash is more expensive than spinning disk, but the development of multi-level cell (MLC) flash, triple-level cell (TLC) NAND flash and 3D NAND flash has lowered the cost. These technologies enable greater flash density without the cost involved in shrinking NAND cells.
MLC flash is slower and less durable than single-level cell (SLC) flash, but companies have developed software that improves its wear level to make MLC acceptable for enterprise applications. SLC flash remains the choice for applications with the highest I/O requirements, however. TLC flash reduces the price more than MLC, although it also comes with performance and durability tradeoffs that can be mitigated with software. Vendor products that support TLC SSDs include the Dell EMC SC Series and Kaminario K2 arrays.
Considerations for buying an all-flash array
Deciding to buy an AFA involves more than simple comparisons of vendor products. An all-flash array that delivers massive performance increases to a specific set of applications may not provide equivalent benefits to other workloads. For example, running virtualized applications in flash with inline data deduplication and compression tends to be more cost-effective than flash that supports streaming media in which unique files are uncompressible.
An all-SSD system will produce smaller variations than that of an HDD array in maximum, minimum and average latencies. This makes flash a good fit for most read-intensive applications.
The tradeoff comes in write amplification, which relates to how an SSD will rewrite data to erase an entire block. Write-intensive workloads require a special algorithm to collect all the writes on the same block of the SSD, thus ensuring the software always writes multiple changes to the same block.
Garbage collection can present a similar issue with SSDs. A flash cell can only withstand a limited number of writes, so wear leveling can be used to increase flash endurance. Most vendors design their all-flash systems to minimize the impact of garbage collection and wear leveling, although users with write-intensive workloads may wish to independently test a vendor's array to determine the best configuration.
Despite paying a higher upfront price for the system, users who buy an AFA may see the cost of storage decline over time. This is tied to an all-flash array's increased CPU utilization, which means an organization will need to buy fewer application servers.
The physical size of an AFA is smaller than that of a disk array, which lowers the rack count. Having fewer racks in a system also reduces the heat generated and the cooling power consumed in the data center.
All-flash array vendors, products and markets
Flash was first introduced as a handful of SSDs in otherwise all-HDD systems with the purpose to create a small flash tier to accelerate a few critical applications. Thus was born the hybrid flash array.
The next phase of evolution arrived with the advent of software that enabled an SSD to serve as a front-end cache for disk storage, extending the benefit of faster performance across all the applications running on the array.
The now-defunct vendor Fusion-io was an early pioneer of fast flash. Launched in 2005, Fusion-io sold Peripheral Component Interface Express (PCIe) cards packed with flash chips. Inserting the PCIe flash cards in server slots enabled a data center to boost the performance of traditional server hardware. Fusion-io was acquired by SanDisk in 2014, which itself was subsequently acquired by Western Digital Corp.
Also breaking ground early was Violin, whose systems -- designed with custom-built silicon -- gained customers quickly, fueling its rise in public markets in 2013. By 2017, Violin was surpassed by all-flash competitors whose arrays integrated sophisticated software data services. After filing for bankruptcy, the vendor was relaunched by private investors as Violin Systems in 2018, with a focus on selling all-flash storage to managed service providers.
All-flash array vendors, such as Pure Storage and XtremIO -- part of Dell EMC -- were among the earliest to incorporate inline compression and data deduplication, which most other vendors now include as a standard feature. Adding deduplication helped give AFAs the opportunity for price parity with storage based on cheaper rotating media.
A sampling of leading all-flash array products includes the following:
- Dell EMC VMAX
- Dell EMC Unity
- Dell EMC XtremIO
- Dell EMC Isilon NAS
- Fujitsu Eternus AF
- Hewlett Packard Enterprise (HPE) 3PAR StoreServ
- HPE Nimble Storage AF series
- Hitachi Vantara Virtual Storage Platform
- Huawei OceanStor
- IBM FlashSystem V9000
- IBM Storwize 5000 and Storwize V7000F
- Kaminario K2
- NetApp All-Flash Fabric-Attached Array (NetApp AFF)
- NetApp SolidFire family -- including NetApp HCI
- Pure Storage FlashArray
- Pure FlashBlade NAS/object storage array
- Tegile Systems T4600 -- bought in 2017 by Western Digital
- Tintri EC Series
Impact on hybrid arrays use cases
Falling flash prices, data growth and integrated data services have increased the appeal of all-flash arrays for many enterprises. This has led to industry speculation that all-flash storage can supplant hybrid arrays, although there remain good reasons to consider using a hybrid storage infrastructure.
HDDs offer predictable performance at a fairly low cost per gigabyte, although they use more power and are slower than flash, resulting in a high cost per IOPS. All-flash arrays also have a lower cost per IOPS, coupled with the advantages of speed and lower power consumption, but they carry a higher upfront acquisition price and per-gigabyte cost.
A hybrid flash array enables enterprises to strike a balance between relatively low cost and balanced performance. Since a hybrid array supports high-capacity disk drives, it offers greater total storage than an AFA.
All-flash NVMe and NVMe over Fabrics
All-flash arrays based on nonvolatile memory express (NVMe) flash technologies represent the next phase of maturation. The NVMe host controller interface speeds data transfer by enabling an application to communicate directly with back-end storage.
NVMe is meant to be a faster alternative to the Small Computer System Interface (SCSI) standard that transfers data between a host and a target device. Development of the NVMe standard is under the auspices of NVM Express Inc., a nonprofit organization comprising more than 100 member technology companies.
The NVMe standard is widely considered to be the eventual successor to the SAS and SATA protocols. NVMe form factors include add-in cards, U.2 2.5-inch and M.2 SSD devices.
Some of the NVMe-based products available include:
- DataDirect Networks Flashscale
- Datrium DVX hybrid system
- HPE Persistent Memory
- Kaminario K2.N
- Micron Accelerated Solutions NVMe reference architecture
- Micron SolidScale NVMe over Fabrics appliances
- Pure Storage FlashArray//X
- Tegile IntelliFlash
A handful of NVMe-flash startups are bringing products to market, as well, including:
- Apeiron Data Systems combines NVMe drives with data services housed in field-programmable gate arrays instead of servers attached to storage arrays.
- E8 Storage E8-D24 NVMe flash arrays replicate snapshots to attached compute servers to reduce management overhead on the array.
- Excelero software-defined storage runs on any x86 server.
- Mangstor MX6300 NVMe over Fabrics (NVMe-oF) storage is branded PCIe NVMe add-in cards on Dell PowerEdge servers.
- Pavilion Data Systems-branded Pavilion Memory Array.
- Vexata VX-100 is based on the software-defined Vexata Active Data Fabric.
Industry experts expect 2018 to usher in more end-to-end, rack-scale flash storage systems based on NVMe-oF. These systems integrate custom NVMe flash modules as a fabric in place of a bunch of NVMe SSDs.
The NVMe-oF transport mechanism enables a long-distance connection between host devices and NVMe storage devices. IBM, Kaminario and Pure Storage have publicly disclosed products to support NVMe-oF, although most storage vendors have pledged support.
All-flash storage arrays in hyper-converged infrastructure
Hyper-converged infrastructure (HCI) systems combine computing, networking, storage and virtualization resources as an integrated appliance. Most hyper-convergence products are designed to use disk as front-end storage, relying on a moderate flash cache layer to accelerate applications or to use as cold storage. For reasons related to performance, most HCI arrays were not traditionally built primarily for flash storage, although that started to change in 2017.
Now the leading HCI vendors sell all-flash versions. Among these vendors are Cisco, Dell EMC, HPE, Nutanix, Pivot3 and Scale Computing. NetApp launched an HCI product in October 2017 built around its SolidFire all-flash storage platform.