This vendor-written tech primer has been edited by Network World to eliminate product promotion, but readers should note it will likely favor the submitter's approach.
Over the past few years mainstream enterprises have been turning to NAND flash storage to boost speed and decrease latency, but some vendors still produce products that inhibit customers from achieving flash's full potential.
Solid-state storage offerings that integrate NAND flash as they would traditional disk systems put data far away from the CPU, often behind an outdated storage controller. No matter how fast the NAND is, this setup creates latency, ensuring the application sees only small improvements in actual throughput.
Let's take a step back and look at the pain of disk storage, the pitfalls of applying conventional architectures to flash, and how to achieve the full potential of NAND flash.
The speed limitations of disk drives compared to CPUs are well known. Less well known are the disk acrobatics administrators have to endure to configure drives for performance. This includes buying expensive Fibre Channel disk drives and configuring them in complex schemes that use only a portion of the drive platter to boost performance, which means adding stacks of disks with largely unused capacity that administrators must monitor for failure (not to mention the costs for power, cooling and space to house the systems).
But even with these acrobatics, disks often struggle to meet required performance levels due to the distance of external disk storage systems from the CPU, as shown in Figure 1. While CPUs and memory operate in microseconds, access to external disk-based systems happens in milliseconds -- a thousandfold difference. Even when disk systems can pull data quickly, getting the data to and from the CPU has a long latency delay causing CPUs to spend a lot of time waiting for data. This negatively impacts application and database performance.
If you consider flash as a new form of media, like tape and disk drives are media, then implementing it the same way you implemented previous media technologies is only a small part of the way forward.
By itself, flash removes the part of the latency bottleneck caused by slow spinning disk drives, but it does nothing to resolve the delay in getting process-critical data to and from the CPU.
Storing data in a flash array puts process-critical data on the wrong side of the storage channel, far away from the server CPU that is processing application and database requests.
The result is a minimal performance gain and, in addition to adding more hardware, organizations must also implement complex and costly storage area network infrastructure, including host bus adapters, switches and monolithic arrays.
But most importantly, these architectures retain the traditional implementations of storage, as well as RAID, and SATA/SAS controllers -- all optimized to spinning drives, not NAND flash silicon. Figure 2 shows the layers still present in this legacy approach.
Increasingly, solid-state vendors have recognized that the key to realizing improved performance is putting flash close to the CPU, and they are creating devices that use PCIe natively, without the inhibitors of outdated translation layers.
BACKGROUND: Flash storage moves closer to CPUs
However, some of these devices hamstring performance by placing the flash under the control of legacy storage implementations of SATA or SAS controllers that were initially designed for disks. These protocols and data handling mechanisms were never intended to operate with NAND flash and do not do any justice to NAND flash capabilities. It's like putting a performance automobile engine into the body of a 25-year-old clunker.
The same thing goes for RAID controllers. Initially designed to aggregate the performance of multiple disks and protect from individual disk failures, conventional RAID mechanisms work well for spinning media. However, these mechanisms do not work well for NAND flash, because they inject too much latency.
The best mechanism to place flash in a server is referred to as native PCIe access, where legacy storage technologies are put aside, and a new cut-through architecture provides the most direct, accessible, and lowest latency path between the NAND flash and the host memory.
Keep in mind that CPUs never read information from storage; instead, everything must pass through system memory first. To assist in the process, native PCIe NAND flash devices present storage to the application or database like a disk drive, but they actually deliver the data to the system memory via Direct Memory Access, or DMA. This guarantees the lowest latency transactions between data storage and CPU processing.
By offering server CPUs unrestricted access to flash, native PCIe implementations increase application and database performance 10x. The difference between this cut-through approach and other solid-state offerings is the improvement to application throughput and not just raw media performance. Data placement in the server without legacy storage protocols allows applications to fully utilize server CPUs by not forcing them to wait for slow access to data, as shown in Figure 3.
Using flash as a disk or a cache
A native PCIe NAND flash device can be used as a disk drive or a caching device. Both provide significant advantages to conventional disk-based systems.
In disk drive mode, a NAND flash PCIe device can store data as if it were a disk drive itself. This is ideal for databases where the entire data set can be placed on one or more PCIe devices. NAND flash PCIe devices can be aggregated with host OS software or built-in volume management functionality, such as Oracle Automatic Storage Management (ASM). Using high- capacity native NAND flash PCIe devices, it is possible to get well over 10 terabytes in a single server -- plenty of capacity to cover a broad market for this approach. Even if the entire data set cannot be placed in flash, most databases allow for the placement of active files such as index files or "hot" tables to be manually placed on a specific data store.
In caching mode, a NAND flash PCIe device can cache frequently accessed data without changing the existing external storage infrastructure. This is ideal where existing subsystem-based data protection and recovery mechanisms are in place.
Caching frequently accessed data locally within each server guarantees the maximum performance for active data while still retaining existing data stores. This combination is ideal for I/O-intensive applications on bare metal or for virtual environments. In many cases virtual environments suffer from inadequate I/O capabilities, or from I/O that can only be achieved at high costs. Caching frequently accessed virtual machine data locally on PCIe flash devices alleviates this pain.
Flash technology brings a lot to the table for speeding up enterprise applications and databases. But when flash is treated as just a new kind of disk drive, businesses miss the mark in delivering on its full potential. Native PCIe approaches that forgo legacy disk protocols and place process-critical data near the CPU to minimize latency deliver on flash's promise to the enterprise.
Fusion-io has pioneered a next-generation storage memory platform that significantly improves the processing capabilities within a data center by relocating process-critical, or "active," data from centralized storage to the server where it is being processed, a methodology referred to as data decentralization. Fusion-io's platform enables enterprises to increase the utilization, performance and efficiency of their data center resources and extract greater value from their information assets.