NetApp announced 8/21/2012 an excellent, necessary, but insufficient set of flash products and messages. The announcement includes:
- Extending the Virtual Tier strategy to include server flash caches;
- Announcing a reseller agreement with Fusion-io hardware, and its IOTurbine and directCache software;
- Announcing Flash Accel caching software that runs in a server or VM and virtual servers(s).
The good news is that NetApp have chosen Fusion-io as a Tier 1 supplier of cache and industry-leading software SKUs, and NetApp has signed a reseller agreement with Fusion-io. The two software products supported are IOTurbine, which supports Windows Server 2003 & 2008; and directCache, which supports Windows 2008 & most Linux distributions.
NetApp has also introduced a product called Flash Accel, software that runs in a server or VM. A NetApp diagram for Flash Accel is shown in Figure 1 below:
Source: NetApp Presentation August 21 2012
The following comments are made by NetApp about Flash Accel in servers:
- Choose your own flash device from a list of qualified PCIe/SSD devices from Fusion-io, FlashSoft, LSI, Micron, STEC & Virident;
- Use for business-application specific performance hot spots with a high read-to-write ratio;
- Enjoy persistent and durable cache across VM / server reboots (presumably across a single VMware physical system);
- Attain intelligent data coherency: block-level invalidation rather than flush entire cache;
- Achieve deep ONTAP integration - ONTAP on the Server. (As discussed later, this statement is misleading).
A close examination of the diagram focuses on the blue control lines and the fact that the total system is inside a single VMware physical server. The key questions are all around how the last two claims are met and where “the master copy of the data” is held. The bottom line is that the master copy is held in the storage array. The cache is read-only with write-though to the storage array. To write to data that is held in read-only cache, that data has to be invalidated, and applications waiting to read that data will have to wait for the acknowledgment to come back to release the lock and for the data to be fetched from the array. The speed difference between the server and the array means that only data with very, very low levels of write or read-only data can be used in the server cache.
It is also clear that Flash Accel is not managed by ONTAP - Flash Accel works completely independently. NetApp highlighted a capability of ONTAP being able to detect that the Flash Accel is out of line with the "master copy of data in ONTAP". There does not appear to be a situation where this error could occur that is not caused by either a malfunction of the data system or a mistake in setup of the data system. If there was such an out-of-line situation, every operations manager, application manager or DBA known to man would bring the system down before more havoc was created, and correct the environment.
The NetApp argument is that caching will allow efficient use of flash; simple installations; self-managing, non-disruptive operations; minimal overhead for data migration; minimal HDD IOs; and highly granular and real time responsiveness. The tag line is “Intelligent, Infinite & Immortal”. The NetApp strategy is applicable for existing cache-friendly applications that will tolerate very large variances of IO response time caused by caching.
However, over time this approach will significantly reduce the available market for NetApp storage arrays. NetApp shared the figures of the amount of flash utilized on its storage (17 Petabytes on 1.2 Exabytes). This means that the average flash penetration is 1.7% of total storage. Typical flash caches are about 3% of total storage. NetApp must be aiming at far higher levels of flash adoption, and the Flash Pool capabilities will not take it to any reasonable target to retain market share.
Both Wikibon and Gartner are predicting a very strong growth of flash-only arrays by 2015. The advantages of flash-only caches for active data are strong, among the most compelling being consistent low variance, very fast IO response time, extreme simplicity of management, and a complete separation of the cost of holding data from the cost of accessing data. The business benefits from eliminating the traditional IO bottleneck are immense. A Wikibon Case Study of Revere shows the potential benefit of allowing greater application integration to be 30% or higher increase in the productivity of an organization. The business value from using 100% persistent storage will overwhelm the potential low levels of saving from slightly cheaper storage.
EMC with the purchase of XtremIO (& perhaps IBM with the purchase of Texas Memory) have understood the direction and are preparing products that will compete with the new comers in the flash-only space. A lot of new ideas are coming from Astute Networks, Kaminario (using Fusion-io technology), NexGen Storage (hybrid* flash/HDD), Nimble (hybrid* flash/HDD), Nimbus Data Systems, Pure Storage, SolidFire, Tintri (hybrid* flash/HDD), Virident, Violin Memory, and Whiptail.
Action Item: Users should be wary of choosing NetApp as a long-term flash strategic partner, while it continues a cache-only strategy. NetApp is taking a path which maximizes short-term ease of deployment, but which puts the master copy of data firmly on disks on arrays. Data management at this level is unsustainable in the long-term and will not be applicable to the new versions of software that will be expecting flash-only arrays. Users should encourage NetApp to take on a broader systems approach to providing a top-down data management framework and true flash-only arrays, either through hiring new talent, acquisition, or a closer collaboration with Fusion-io.
Footnotes: Note: * The definition of hybrid is that there is no under-provisioning of IO access to flash storage (all flash storage can operate all the time at its rated speed. Traditional arrays, where the disk trays are fully populated with flash drives but the disk trays are not modified to avoid under-provisioning of IO access, do not meet the definition of hybrid storage)
