Letting Cache Acceleration Cards Do The Heavy Lifiting

Up until now there has not been a great deal of intelligence around SSD Cache cards and flash arrays because they have primarily been configrued as DAS (Direct Attach Storage). By moving read intensive workload up to the server off of a storage array, both individual application performance as well as overall storage performance can be enhanced. There are great benefits to using SSD Cache cards in new ways yet before exploring new capabilities it is important to remember the history of the products.

The biggest problem with hard drives either local or SAN based is that they have not been able to keep up with Moore’s Law of Transistor Density. In 1965 Gordon Moore, a co-founder of Intel, made the observation that the number of components in integrated circuits doubled every year – he later (in 1975) adjusted that prediction to doubling every two years. So, system processors (CPUs), memory (DRAM), system busses, and hard drive capacity have been doubling in speed every two years, but hard drives performance has stagnated because of mechanical limitations. (mostly heat, stability, and signaling reliability from increasing spindle speeds) This effectively limits individual hard drives to 180 IOPs or 45MB/sec under typical random workloads depending on block sizes.

The next challenge is that in an effort to consolidate storage, increase the number of spindles, availability and efficiency we have pulled the storage out of our servers and placed that data on SAN arrays. There is tremendous benefit to this, however doing this introduces new considerations. The network bandwidth is 1/10th of the system bus interconnect (8Gb FC = 1GB/sec vs PCIe 3.0 x16 = 16GB/sec). An array may have 8 or 16 front-end connections yielding and aggregate of 8-16GB/sec where a single PCIe slot has the same amount of bandwidth. The difference is the array and multiple servers share its resources and each can potentially impact the other.

Cache acceleration cards address both the mechanical limitations of hard drives and the shared-resource conflict of storage networks for a specific subset of data. These cards utilize NAND flash (either SLC or MLC, but more on that later) memory packaged on a PCIe card with an interface controller to provide high bandwidth and throughput for read intensive workloads on small datasets of ephemeral data.

[framed_box bgColor=”#F0F0F0″ textColor=”undefined” rounded=”true”] I realize there was a lot of qualification statements there so lets break it down…

  • Why read intensive? As compared to SLC NAND flash, MLC NAND flash has a much higher write penalty making writes more costly in terms of time and overall life expectancy of a drive/card.
  •  Why small datasets? Most Cache acceleration cards are fairly small in comparison to hard drives. The largest top out at ~3TB (typical sizes are 300-700GB) and the cost per GB is much much higher than comparable hard drive storage.
  •  Why ephemeral data and what does that mean? Ephemeral data is data that is temporary, transient, or in process. Things like page files, SQL server TEMPDB, or spool directories.
[/framed_box] Cache acceleration cards address the shared-resource conflict by pulling resource intense activities back onto the server and off of the SAN arrays. How this is accomplished is the key differentiator of the products available today.

SSD Caching , EMC, VFCache, FusionIO, VMWare, SLC, MLC NAND Flash, Gordon Moore, Intel, processors, CPU's, DRAM

FusionIO is one of the companies that has made a name for themselves early in the enterprise PCI and PCIe Flash cache acceleration market. Their solutions have been primarily DAS(Direct Attach Storage) solutions based on SLC and MLC NAND Flash. In early 2011 FusionIO released write-through caching to their SSD cards with their acquisition of ioTurbine software to accelerate VMWare guest performance. More recently – Mid-2012 – FusionIO released their ION enterprise flash array – which consists of a chassis containing several of their PCIe cards. They leverage RAID protection across these cards for availability. Available interconnects include 8Gb FC and Infiniband. EMC release VFCache in 2012 and has subsequently released two additional updates.

The EMC VFCache is a re-packaged Micron P320h or LSI WarpDrive PCIe SSD with a write-through caching driver targeted primarily at read intensive workloads. In the subsequent releases they have enhanced VMWare functionality and added the ability to run in “split-card” mode with half the card utilized for read caching and the other half as DAS. EMC’s worst kept secret is their “Project Thunder” release of the XTremIO acquisition. “Project Thunder” is an all SSD array that will support both read and write workloads similar to the FusionIO ION array.

SSD Caching solutions are an extremely powerful solution to very specific workloads. By moving read intensive workload up to the server off of a storage array, both individual application performance as well as overall storage performance can be enhanced. The key to determining whether or not these tools will help is careful analysis around reads vs writes, and the locality of reference of active data. If random write performance is required consider SLC based cards or caching arrays over MLC.


Images courtsey of “the register” and “IRRI images