Facebook engineers say bigger hard disk drives is making one critical metric far, far worse
- Hard drives are getting larger, but performance is a problem in data centers
- Meta sees QLC flash as a solution, bridging the gap between HDDs and TLC
- With 512TB QLC SSDs planned, Meta is shifting away from E1.S and E3
Hard disk drives are still the go-to storage choice for most data centers, mainly because they are cheaper and use less power than TLC flash.
But while HDDs are getting ever larger, their input/output performance hasn’t kept up. As a result, bandwidth per terabyte has been shrinking, forcing data centers to move frequently accessed data to TLC flash or overprovision storage.
In a recent Facebook Engineering blog post, Meta outlined its approach to integrating QLC flash as a solution to these challenges. QLC flash has existed since 2009, but adoption has been slow due to limited capacity, high cost, and lower write endurance.
You may like
QLC SSDs are the future
Meta explained it sees QLC as the middle ground between HDDs and TLC SSDs. The technology delivers enough performance for workloads that previously relied on 16TB and 20TB HDDs and supports large batch I/O tasks that need more than HDDs can offer. Since most power consumption in flash storage comes from writes, QLC’s lower write activity makes it a feasible, energy-efficient option.
While QLC is currently cheaper than TLC, it is not yet at a price point for widespread deployment. The power savings make it an attractive option for certain workloads however, and recent advancements are making the flash format more practical. The arrival of the 2Tb QLC NAND die and the widespread use of 32-die stacks are driving higher storage density and QLC SSDs are expected to scale faster than TLC. This will ultimately help to improve server efficiency while lowering costs.
Meta is planning for QLC SSDs to reach capacities as large as 512TB and sees E1.S and E3 form factors as dead on arrival for its needs. Instead, it intends to focus on U.2-15mm and Pure Storage’s DirectFlash Module for scaling QLC in its infrastructure.
Meta is also refining its storage software to handle the high-density nature of QLC. Since QLC SSDs have a large gap between read and write speeds, proper scheduling is needed to maintain performance.
As QLC technology advances, Meta expects its role in data centers to grow, a view that’s shared by StorageReview.
The site’s Brian Beeler writes, “Our testing has reinforced QLC’s role in AI storage. A recent deep dive explored how Solidigm’s D5-P5336 QLC SSD performs in checkpointing workloads for AI model training. These tests showed that while TLC SSDs still lead in write-heavy environments, QLC SSDs hold their own regarding capacity, efficiency, and read performance – making them a strong option for AI pipelines.“
You might also like
Hard drives are getting larger, but performance is a problem in data centers Meta sees QLC flash as a solution, bridging the gap between HDDs and TLC With 512TB QLC SSDs planned, Meta is shifting away from E1.S and E3 Hard disk drives are still the go-to storage choice for…
