In the early days of storage, IO was all about disc. The more spinning disc you placed into a lun, the more redundancy, IO, and the less rebuild time in the case of disc failure required. This turned out to be a solid but expensive approach. When faster spinning disc became the norm, 10k or even 15k discs added IO to the platform as well. You could reasonably anticipate 170 or so IOPS out of a standard 10K (10,000 Revolutions per minute) spinning disc. So, theoretically, a LUN with 6x10K discs could achieve 1000 relatively easily. In its day, that was a solid performance metric. Would give ample redundancy, and would not take such a large hit in terms of failure disc rebuild. But, that’s really still not very fast, now is it?
In terms of, for example, VDI, this lun, at a peak speed per desktop of roughly 35 Iops read and write, this’ll support at most 30 instances. Once can easily see how deploying a true VDI environment on spinning disc became a significant challenge.
Along came Solid State disc. In many cases, the IO from these can exceed 30,000 IOps per disc, and in the case of truly enterprise class, high-performance SSD, it is not uncommon to see double that speed or more. No longer are we constrained by adding disc to a disc group in order to gain efficiency at the expense of space. I’ll grant you that these discs are often quite a bit more expensive, but with more intelligence packed onto the controller, these discs have additional feature including longevity and erasure coding unlike consumer discs.
Imagine the difference of a 3-disc group, (which provides ample redundancy in case of failure) each at 16tb ( the newest Samsung SSD’s max capacity [ http://bit.ly/1poOqiT/ ] totaling, in raid 5, approximately 36Tb of usable space, with failure abilities, and throughput at, in excess of 40,000 iops (this is rated at 18,000 IOPS per disc). Significant differences in terms of speed, energy consumption, configurations, etc. abound. Not to mention, the tweaks necessary to optimize the individual desktop performance statistics became far less difficult to optimize.
It would be conceivable to place thousands of desktops into only 2 rack units of storage.
These newer SSD formats change the game so significantly that the issues in terms of latency become more concrete than the discs themselves. For example, SAS connectivity versus SATA, Fibre versus iSCSI, Block versus File versus Object, these things become more the choices that need to be made than quantity of disc per group. But again, outside of HPC (High Performance Computing) Seismic, Biomedical Engineering, Big Data, etc, the Disc IO per second is no longer the issue, is it? I think not.
Today, when I discuss storage with my customers, while my approach is pretty much identical. My initial questions are always “What do you hope to accomplish today, and into the future?” but the considerations regarding power, cooling, connectivity, speed, etc. are easier to accommodate. We can think in terms of cloud, replications and integrations in different and much more interesting ways. In terms of costs per gigabyte, with the inclusion of real, non disruptive data integrity checking, erasure coding, deduplication and compression, these costs have achieved a level of parity previously unanticipated.
Today’s issues are more around how easily managed, how expensive, and how scalable is it?
We truly live in a different world of storage. Software Defined, with the inclusion of commodity equipment is a new paradigm. The game’s changed, the players are changing and the considerations are brand new.