The nature of Flash Storage

We began today discussing the nature of flash storage, specifically, how it is built from a MOSFET-like semiconductor in strings of pages forming an array known as a block, aggregated into a flash array. From there we talked about the flash translation layer (FTL) and garbage collection.

This part of the discussion followed NAND Flash memory Organization and Operations (Nvotny, Kadlec, and Kuchta).

In addition to the usual discussion of the need for write-leveling, we also discussed the potential need for read-leveling, Read leveling for flash storage systems (Liu; Chang YM; Chang YH)

Impact of Flash Memory on System Design

As Flash Memory has becoming faster, more scalable and less expensive, it has evolved from an accessory to consumer products, such as phones and cameras; to an end-user storage solution, e.g. instead of hard drives for laptops; to a part of the enterprise and data center storage hierarchy, e.hg. as a cache above disk and for persistent journaling.

We also discussed that as these solutions scale horizontally, they can quickly saturate the bandwidth of the bus technology commmonly used for storage. And, beyond that, it is likely to become increasingly difficulty for processor burst cycles to dominate I/O in many applications. Goingf forward, this may call into question the existing interrupt-driven I/O model as well as the model in which persistent storage is incorporated as second-class storage vs addressable memory. This may be especially true of, for example, systems that use capacitor backed DRAM as buffers to write to SSD storage.

A good read related to our discussion is this perspective paper, Non-volatile Storage: Implications of the Datacenter's shifting center (Nanavati, Schwarzkopf, Wires, Warfield).

Flash Memory vs Disk

As we discussed last class, the long term trend w.r.t. disk storage is that disks storage is getting less expensive much more rapidly than it is getting faster. It remains much less expensive than Flash storage. And, it seems that it will be a very long time before SSD catches up.

For end users, we already largely prefer SSD storage, despite the modest increase in cost. Most of our laptops have Flash not hard drives. But, when it comes to those who play with large volumes of data -- cost matters. Hard drives will seemingly remain the backbone of big data long after they are forgotten by consumers. They'll be augmented by flash upstream. But, until flash becomes cheaper per byte -- hard drives are here to stay. At least underneath the cloud.