The term Dynamic Tiering has really been abused in the network storage industry lately. Everyone talks about tiering in one form or another. Vendors that do not have currently shipping products talk about futures and those vendors that have shipping products do not actually disclose how their tiering works, which applications benefit, nor by how much.
Unfortunately, one of the giants in the industry has even further muddied the waters. According to a TechTarget article last week regarding NetApp’s earnings call, NetApp’s CEO said that “the whole concept of tiered storage is going to go away.” Presumably, this refers to EMC’s Fast technology, since one industry Goliath always needs to beat up on another Goliath. The unfortunate thing for NetApp, is that on the same call, they completely reversed their position on tiering when they touted the success of their form of tiering, which includes their SSD-based performance accelerator. Tiering is dead, long live tiering!
The simple truth is that no single technology has ever proven to be the panacea of data storage. SATA drives have the lowest cost per bit and are great for archival storage. FC or SAS drives offer a compromise of performance and capacity, excelling at large block accesses like those found in large sequentially read files. Solid State Devices, based on Flash, offer unsurpassed performance for random reads and small block sizes.
More importantly, the differences from one technology to another are measured by orders of magnitude, not by mere percentage improvements. Because of this, a solution that can leverage the strengths of all the technologies is guaranteed to out-perform and cost less than a solution that only uses one or two.
Rather than predicting which storage tiers will win and the capacity of those tiers in a solution, the important information needed to judge a tiering solution is how the tiers are used. Most vendors are completely silent on this. Here are three examples of the more egregious mis-steps in dynamic tiering.
The first mis-step is tiering at too large a level of granularity. Consider a solution that tiers at the volume level. If a few files in a volume become active, the entire volume will need to be promoted to a more expensive tier to get the performance needed for the few files. This results in cost inefficiency as extra data is promoted to the expensive storage and performance inefficiency as the entire volume consumes read/write bandwidth of both tiers that are involved in the promotion of the volume.
The second mis-step is not tiering frequently enough. Several vendors have proposed tiering schedules that are measured in terms of days. This is crazy. Consider the file that I am editing for this blog. I might work on this file for a few days and then rarely, if ever, look at the source file again once the blog is posted. If activity is measure across days, by the time this source file is promoted, it should be archived.
The third mis-step is not using the correct media. Most vendors actually completely avoid this question and require the administrator to set policies. In those instances where the vendor does decide, frequently wrong media is chosen.
An example of this third mis-step is to examine the two stage architecture promoted on NTAP’s earnings call – SSD & SATA. In their architecture the SSD-based performance accelerator is apparently only used for read data. All write data is sent to SATA storage. This is terribly inefficient and is even proven in their SPECsfs®08 posting. To achieve the same performance in a NTAP 3160 with the accelerator module, they required almost twice the number of SATA disks than when they run the same benchmark using FC disks (96 SATA disks versus 56 FC disks). Since the SATA disks have over 3x the capacity of the FC disks, they deployed over 6x as much capacity to store the same amount of data. This over-provisioning is a result of not tiering the media properly and is extremely inefficient in terms of space, power and equipment costs.
What is clear from all of the press on “Dynamic Tiering”, is that the term is both extremely overused and misunderstood. Because of the orders of magnitude differences in storage media costs and performance, data storage solutions can clearly benefit from tiering if executed properly.
