TLC, QLC, SCM:
Making sense of Alphabet soup to determine which is right for you

Customers typically ask us what type of flash Pavilion uses in its Hyperparallel Flash Array (HFA). Given that there are a number of different types of memory available for use in all flash arrays (AFAs), the key customer question should be which one will best meet their requirements for performance, capacity, and lifespan, as well as the resulting cost of the array.

TLC:
Pavilion uses TLC SSDs in its HFA because we believe TLC provides the best outcomes for customers. As its name implies, Triple Layer Cell (TLC) SSDs, write three bits to each memory cell. TLC drives offer an optimum combination of performance, density, and longevity.

QLC:
So what about QLC SSDs? Quad level cell (QLC) SSDs write four bits to each memory cell and it can store more data in the same area, which often makes QLC more suitable for high density environments. But, the drawback of QLC is its limited write endurance.

Write endurance is a measure of the number of times that a program/erase (P/E) cycle can be applied to a particular cell before it becomes unreliable. P/E cycles can be equated to read/write operations.

Traditional hard drives can be written to and read from until one of the physical moving parts in the drive fails. SSDs have no moving parts, but are governed by an inherent characteristic of the media. Once a given cell has been written to a certain number of times, the cell will become unstable and data can be lost. A characteristic of NAND memory is that it can be read from far more times than it can be written to, so the lifespan of an SSD is measured in terms of write endurance.

Endurance is measured by drive writes per day (DWPD), which is the number of times that a drive can be fully written to in a single day, every day for the given life of the drive. So, a 1 TB SSD with a write endurance of 1DWPD and a 10 year life can have 1 TB of data written to it every day for 10 years. Specifications will vary across manufacturers, but in general TLC SSD write endurance is usually between 1DWPD and 3DWPD. QLC SSDs are about 0.1DWPD. This makes QLC SSDs impractical for write intensive workloads, but good for read heavy environments.

SCM:
Storage Class Memory (SCM) is a different class of memory than TLC and QLC and it sits between DRAM and NAND flash in terms of performance. One of the more common types of SCM is 3D XPoint, which is often marketed under the brand name Optane. This memory is an order of magnitude faster than NAND flash, but that performance comes at a cost. SCM may be faster than NAND flash, but it can also be an order of magnitude more expensive.

How Manufacturers Decide Which to Offer:
Some storage manufacturers try to balance performance and cost by using a combination of SCM and QLC memory. In theory this alphabet soup sounds good; however, in practical application it is almost impossible to get “just right”. (How often can you really spell a sentence let alone a five letter word out of a can of ABC’s?)

The issue is trying to balance the amount of high cost SCM for write operations and QLC for read operations. If the data set is too large, data will be offloaded to the QLC for writes, massively impacting performance and the life of the array. Conversely, if the data set is too small, then a very expensive array will be underutilized. Since data sets, by their very nature, change over time, getting the right mix of SCM and QLC is almost impossible.

Recognizing these tradeoffs, Pavilion uses TLC SSDs in its HFA. As a result, it is able to deliver significantly greater performance than systems that use SCM as their performance is tiered. So if SCM is an order of magnitude faster than TLC, how can the Pavilion HFA be so much more performant?

The Secret Sauce:
The Pavilion HFA is significantly faster than its industry peers because the performance constraint in most AFAs is not the type of memory used, it is the array architecture.

Typical AFAs are built using a dual controller architecture, which is a legacy design from the hard drive era and is the true bottleneck. The Pavilion HFA uses a modern network switch design which unlocks the parallel performance potential of NAND. The result is that a single HFA can deliver up to 120GB/s, over 20M IOPs, and as little as 25µs of latency measured from the host, from each 4RU system. That performance scales linearly. One HFA delivers 120GB/s in 4RU, two HFAs deliver 240GB/s in 8RU, and so on.

The Pavilion HFA is also able to deliver this outcome for less than a typical AFA. While using high cost SCM as a storage tier might deliver impressive results for small data sets, in practice those systems cannot match either the performance or the affordability of The Pavilion HFA.