IT buyers must exercise good judgment in choosing between a read-intensive, mixed-use and write-intensive SSD.
When they first hit the market, solid-state drives (SSDs) tended to wear out relatively quickly. The issue was the finite number of writes each cell of the SSD was able to endure. Write wear, as this issue was called, limited SSD replacement of hard drives. The ensuing decade, however, has seen the problem understood and overcome, and more SSD use cases have emerged.
Apart from substantial improvements to electronics and materials, one major fix for write wear is to overprovision the capacity of the drive, hiding a portion of that raw capacity as replacement blocks. For example, a drive might appear to have 1 TB of capacity, but, in reality, it may have 1.5 TB, with 0.5 TB hidden. This overprovisioning adds to the price of the drive.
Internally, drives don't distinguish between real and replacement blocks. Whenever data is written, its block address is translated to a new physical location in flash. So, as the drive fills up, data spills over into the replacement blocks and the internal drive software recovers the old data blocks at some later time.
The larger the replacement block portion, the longer write wear time the drive offers. Today's market characterizes drives by their wear endurance patterns. Vendor terminology may differ, but, in general, any given SSD family has a read-intensive, mixed-use and write-intensive SSD.
How read- and write-intensive SSDs differ
Read-intensive drives use lower levels of overprovisioning and typically also have less durable flash cells, such as triple-level cells (TLC) that can store more bits, but with a lower write wear endurance. These are the cheapest drives.
Mixed-use drives are more durable, with larger replacement block pools. Depending on their price and capacity, they may use multi-level cells -- two bits per cell -- or TLC flash -- three bits per cell. A write-intensive SSD, the most expensive, can have large replacement pools, which may limit the maximum capacity of the drive.
Buyers must also factor in a drive's performance and the choice of Serial Advanced Technology Attachment (SATA) or NVMe interfaces. In traditional server farms, it isn't too difficult to figure out the write/read ratios of apps in the mix on any given set of storage.
Translating the numbers from HDD to SSD requires some care, however. Apps can speed up dramatically when they move from an HDD with 150 IOPS to an SSD with 50,000 IOPS.
The objective is to estimate the equivalent total drive writes (TDWs) per day that will hit the new SSDs. If the number is around 0.1 TDWs per day, that is likely a read-intensive use case. Ten TDWs per day fit into the write-intensive SSD range, and mixed-use drives sit in the middle.
Performance of the drives is a major factor. NVMe drives may deliver as much as 10 times the IOPS of a SATA drive of the same capacity, and more work means more TDWs per day.
Read- and write-intensive SSD use cases
It's worth looking at some SSD use cases. At the read-intensive end of the range are web and media servers. While these are still typically built with cheap SATA hard drives, the lowering of SSD prices will make SSD much more attractive. High-core edge CPUs hitting the market will accelerate the economic attractiveness of read-intensive SSDs through 2018, too.
Most databases, at the other extreme, are more write-intensive and typically build on NVMe for performance, requiring either a mixed-use or write-intensive SSD. Life gets complicated if a database is prone to having some data that is much more active than other records.
In these cases, use combinations of write-intensive and read-intensive SSDs, with hot data on the write-intensive SSD. Software to auto-tier data between the fast and slow layers is readily available.
In general, data center administrators can get an idea of how other apps operate in terms of I/O, but they need more tools to measure TDW numbers and make it easier to characterize drive usage. Storage analytics tools are becoming more comprehensive and powerful, and much of the data gathering and analysis will be fully automated within three years.