Almost time for SSD flash storage to be used with object storage
As the price of flash SSD storage continues to fall, it becomes more attractive for object storage, which otherwise counts on big, slow and cheap HDDs.
Object storage has become a popular choice for nearline storage, cloud storage, Hadoop storage and even NoSQL storage. The primary media type used for object storage, however, isn't SSD flash storage, but has been and continues to be 3.5-inch, 7,200 rpm fat HDDs. The logic behind this is that object storage is mainly secondary storage. Performance is not the driving factor. Scalability, reliability and durability at a very low cost are the principal factors in object storage sales.
But what if object storage performance could increase by as much as 100 times what it delivers today and with the same or better durability and a denser scalability footprint at the same or lower cost? Based on publicly released roadmaps for read-optimized SSD flash storage and fat HDD, that may indeed be the case by the end of 2016.
It starts with the type of media used in object storage. Since most of the data stored as object storage is infrequently accessed -- it is secondary data after all -- the biggest factors for media are cost and density, not object storage performance. That's why, today, the media is fat drives ranging in raw capacity from 4 TB to 10 TB per HDD, which has the lowest cost per terabyte in the smallest footprint.
Read-optimized SSD flash storage is designed for the same type of data, but in a much smaller 2.5-inch footprint, which takes up 40% of the volume and 10% of the weight of 3.5-inch HDDs. Read-optimized SSDs are designed for approximately 10% writes and 90% reads. The new multilayered 3D or vertical triple-level cell (TLC) NAND is ideal for read-optimized SSDs. The current highest capacity 2.5-inch, read-optimized SSD on the market at the end of 2015 was approximately 4 TB raw. That's changing ... fast.
In Q1 2016, Samsung started shipping a 32-layer, 16 TB, read-optimized, 2.5-inch (SAS or SATA) flash SSD based on 256 GB TLC NAND chips. By Q3, several vendors will be shipping equivalent products. By Q4 2016, expect both 24 TB and 32 TB, read-optimized, 2.5-inch flash SSDs. Intel, Micron, Samsung and Toshiba predict 48 TB and 96 TB, read-optimized, 2.5-inch SSD flash storage shipping in 2017. By 2018, they all expect the gains to accelerate with predictions of 128 TB and 256 TB, 2.5-inch, read-optimized flash SSDs. Compare that with the fat, 3.5-inch nearline drive capacities. Today, the maximum capacity HDD is 10 TB. Seagate and Western Digital predict that will increase to 20 TB by 2020. If they can speed up development, they might -- and this is a low probability -- be able to get that to 40 TB by 2020. Either way, the capacity gap is huge and rapidly growing.
How about reliability?
A recent study of production flash SSDs in the field by the University of Toronto and Google revealed some eyebrow-raising information:
The raw bit error rates (RBER) grow at a much slower rate with wear-out than the exponential rate commonly assumed and, more importantly, they are not predictive of uncorrectable errors or other error modes. Compared to traditional HDDs, flash SSDs have a significantly lower replacement rate in the field; however, they have a higher rate of uncorrectable errors.
Gartner and other analysts are predicting that the read-optimized flash SSD and nearline fat HDD cost-crossover point -- calculating acquisition cost per terabyte -- will occur in 2016.
This means these high-capacity, read-optimized flash SSDs are ideal for object storage. Here's why: SSD program-erase (PE) blocks are far more likely to fail or produce a nonrecoverable bit error on writes because of the very nature of flash NAND. Reads seldom produce a nonrecoverable bit error today, although it is possible with read-disturb situations, which are extremely rare.
Object storage erasure codes essentially write objects as multiple subobjects on different PE blocks, drives and nodes. When a subobject has a nonrecoverable bit error, the object store simply writes it somewhere else and the SSD retires that PE block. It then draws from the overprovisioned stock in the SSD to maintain capacity. This makes read-optimized SSD flash storage quite suitable for object storage.
This is all good, but what about that 800-pound gorilla in the room known as cost? Isn't SSD flash storage -- even read-optimized flash SSDs -- a lot more expensive than HDDs? It used to be. The crossover inflection point between SSDs and high-performance HDDs occurred in 2015. Gartner and other analysts are predicting that the read-optimized flash SSD and nearline fat HDD cost-crossover point -- calculating acquisition cost per terabyte -- will occur in 2016. Keep in mind that read-optimized flash SSD power and cooling costs per raw terabyte average approximately 10%. Add in the data center footprint savings and the savings from not having to use reinforced floor tiles to handle the weight of highly dense HDD racks, and that cost-crossover point is likely to occur before the end of 2016.
A system using read-optimized flash SSDs will have greater object storage performance -- up to two to three times faster -- take up less rack and floor space, and ultimately cost less than today's object storage implementations. The good news for object storage users today is that object storage upgrades are simpler than traditional storage and can occur online one node at a time and without data migration.
