Less than a month ago I wrote a post titled “Two-Stage NAS Storage: Data Delivery and Data Management”. In that post, I wrote that a dynamically tiered NAS architecture only works if the implementation allows the relative capacities of the tiered media to be scaled independently. I reviewed Avere Systems FXT product series, which consists of a dynamically tiered architecture and is based on a two-stage implementation that allows independent scaling of the performance and capacity tiers.
To illustrate independent performance scaling, Avere posted SPECsfs2008_nfs performance numbers for one-node, two-node and six-node FXT clusters and showed linear scaling of the benchmark results.
Since my original post, there has been much press about dynamic tiering. In every case, the vendors claiming dynamic tiering seem to have missed the point. What they are missing is the ability to scale performance tiers independently from capacity tiers. Consider one vendor, NetApp, and their recent SPECsfs2008 postings as an example. In the past few months, NetApp has posted three more server configurations to SPECsfs2008, bringing their total postings to seven. Each of the seven configurations is unique and there is no way to scale non-disruptively from one configuration to another.
For NetApp, performance scaling requires the addition of PAM cards (DRAM or SSD modules designed to accelerate reads), until the chassis runs out of PAM slots or hits a CPU bottleneck, and then the wholesale replacement of the server. The only way to scale non-disruptively is to add disk drives, which scales both performance and capacity at the same time. This results in huge inefficiencies, unless the target application requires performance and capacity in exactly the ratio provided by the added disk drives.
In the Avere FXT postings, one common architecture is represented in all postings: one-node, two-node and six-node FXT clusters to achieve the performance target and a single Mass server to achieve the required capacity. Performance scales linearly among the configurations by adding FXT nodes – nodes are added and join the cluster automatically and on-line. Capacity is scaled among the configurations by adding high-density disk drives to the Mass. This high level of configurability is where Avere’s two-stage implementation of dynamic tiering really excels. It allows the deployment to scale performance and capacity independently to match the exact requirements of the application.
To illustrate the efficiency of the Avere two-stage implementation consider the performance level achieved on the SPECsfs2008 benchmark divided by the number of disk drives used in the configuration, or Ops/disk. This is an excellent metric of the efficiency of a solution as it indicates the amount of equipment needed to achieve a given performance and capacity level.
This also happens to be the subject of a recent blog post by NetApp. What NetApp’s blog fails to indicate is that the Avere SPECsfs2008 postings achieve the highest number of Ops/drive of any vendor (see Figure 1).
For a given performance level and capacity requirement, Avere’s solution uses the least amount of equipment, while the application sees the exact same performance and capacity provided by the NAS server. Avere achieves this by using the highest density disk drives for capacity and faster media tiers to deliver performance which allows the system to scale either attribute independently.
