4 QLC workloads and why they're a good fit for QLC NAND flash

What limits QLC workloads

Enterprise QLC offers a promising replacement for low-performing HDDs deployed extensively in data centers. According to Micron, its QLC drives deliver more than 450 times the read IOPS than HDDs running the same workload. QLC drives also require less space, consume less energy and put SSD prices on par with HDD storage.

A flash drive based on NAND technologies is a type of nonvolatile memory made up of one or more flash chips. Each chip contains dies, each die contains planes and each plane is divided into blocks. The blocks are then divided into pages and the pages into cells, where the data bits are stored. Data is read and written at the page level, but erased at the block level, resulting in a complex write-erase process.

Each drive includes a controller that manages data operations and addresses issues that come with writing and modifying data. For example, the controller is responsible for operations such as wear leveling, garbage collection, bad-block mapping, error code correction and data encryption. Wear leveling and garbage collection are particularly important because they help to extend the drive's endurance, or lifespan, by mitigating some of the effects of the write-erase process.

Several factors contribute to the endurance of a NAND flash drive. At the top of the list is the number of program/erase (P/E) cycles performed during the drive's life. A drive can support only a limited number of P/E cycles before it fails.

One of the biggest factors to impact the number of P/E cycles is write amplification, which results from data modifications that write more data than the amount being modified. This happens because data is written at the page level but deleted at the block level. To complicate matters, a data modification requires data be erased before it can be written, which means an entire block must be erased and rewritten no matter how small the change. Wear leveling and garbage collection, along with over-provisioning, can help address some of these challenges, but the drive is still subject to a limited number of P/E cycles.

QLC NAND flash storage

The number of bits per cells can have a direct impact on how many P/E cycles a drive will support over its lifetime. Although estimates vary, you can safely assume that the more bits per cell, the lower the number of P/E cycles. For example, single-level cell NAND supports around 100,000 P/E cycles, whereas multi-level cell (MLC) NAND supports between 3,000 and 10,000. (MLC refers to drives with two bits per cell.) Estimates for TLC NAND come in around 1,000 to 3,000, and QLC at less than 1,000, with some estimates closer to 100.

The reason higher densities translate to fewer P/E cycles is related to how voltages are applied at the cell level during write operations. Although the process itself is fairly involved, what's important to understand is each write operation causes slight wear to the cell, and the more bits packed into the cell, the greater the wear. For this reason, QLC NAND flash is much more susceptible to wear than the other types of NAND, resulting in fewer P/E cycles.

The right QLC workloads

Because of these limitations, QLC is a poor fit for write-heavy workloads that are constantly adding and updating data. The fewer the write operations, the longer the drive will last. Read-intensive workloads in which writes represent only a small portion of the operations work best with QLC NAND flash. In fact, Micron recommends that less than 10% of a workload should be writes.

Regardless of how QLC evolves, you should always be thinking about the applications you're supporting and the storage they require to deliver their workloads as efficiently and reliably as possible.

Fortunately, a number of today's workloads fit this model. For example, analytical applications that support AI, machine learning, deep learning and other forms of AI are excellent candidates for QLC storage. The data in these scenarios is typically written once and then used to carry out sophisticated analytics, which requires fast access to storage when modeling, training and aggregating massive amounts of data.

Archiving data is another potential QLC workload, especially if an organization uses that data for AI-based analytics or business intelligence applications that mine data to provide stakeholders with near real-time information for quick insights and decision-making. Large data centers that support streaming media operations might also use QLC to deliver audio or video services faster and more efficiently to their customers.

Just about any data store that supports read-intensive applications is a good candidate for QLC. For example, a NoSQL database full of rich data and metadata could use QLC NAND flash drives to boost application performance. As with any scenario, the key is in understanding the extent to which your applications write data compared to how much they read.

Ideal fit vs. money pit

As QLC technologies mature, they'll undoubtedly support a greater number of P/E cycles, expanding the type of workloads that could benefit from QLC flash. Already, we've seen a jump in P/E estimates from less than 100 to nearly 1,000. Even so, QLC will likely continue to be best suited to read-intensive workloads for some time to come.

Regardless of how QLC evolves, you should always be thinking about the applications you're supporting and the storage they require to deliver their workloads as efficiently and reliably as possible. In some QLC workloads, the technology could be an ideal fit, but in others, a costly waste of money.