How does an SLC SSD compare with a DRAM SSD in terms of IOPS? And how do they compare cost-wise? Also, how do you determine if you need the performance of a DRAM solution?
Dynamic RAM (DRAM) SSDs IOPS will typically be about 10 to 12 times greater per GB than those of single-level cell (SLC) SSDs on reads. Write IOPS are astronomically better, especially when an SLC write has to erase data first before it can write.
Comparing costs is a bit more complicated since DRAM SSDs use two to three times more power and cooling than SLC SSDs do, while requiring some form of battery backup or capacitor to allow data to be destaged to disk or SSD during a power failure; otherwise, the data is lost. Generally, DRAM SSD total cost of ownership ranges from four to 10 times that of SLC SSD.
Determining whether a DRAM solution is the way to go requires analyzing the end-to-end ecosystem data path in which it will reside. Just because an application now has a lightning-fast storage solution does not mean there will be a commensurate increase in application performance. It depends on where the performance bottlenecks are and whether the increased IOPS can be effectively utilized.
If the DRAM pushes the bottleneck to the CPU (i.e., the CPU is oversubscribed and can't use the DRAM IOPS), the storage controller (same issue as the CPU), or network IO (it doesn't matter how many IOPS the DRAM can provide if the network can't move them fast enough), then it's probably not a great fit. This is analogous to putting a Porsche engine in a Smart Car. Even if the engine fits, it is highly unlikely the car would be able to effectively utilize it.
On the other hand, if the storage is the primary bottleneck and flash SSD IOPS are just not enough, a DRAM solution is probably appropriate. It comes down to a cost-benefit analysis. If DRAM provides incrementally better results than the costs, it's a good fit.
Dig Deeper on Flash memory and storage
Related Q&A from Marc Staimer
Network File System and Common Internet File System/Server Message Block were designed to work with any operating system, but NFS remains dominant in... Continue Reading
Latency in object stores that stems from a large amount of metadata means the technology is better suited to non-transactional data. Continue Reading
Eventual consistency in object stores can be an issue because object storage is spread over many nodes and up-to-date data may not always be ... Continue Reading