Dell EMC announced that it will soon add Optane-based storage to its PowerMAX arrays, and that PowerMAX will use Optane as a storage tier, not “just” cache. This statement implies using Optane as a storage tier is superior to using it as a cache. But is it?
PowerMAX will use Storage Class Memory as Tier in All-NVMe System
Some people criticized Dell EMC for taking an all-NVMe approach–and therefore eliminating hybrid (flash memory plus HDD) configurations. Yet the all-NVMe decision gave the engineers an opportunity to architect PowerMAX around the inherent parallelism of NVMe. Dell EMC’s design imperative for the PowerMAX is performance over efficiency. And it does perform:
- 290 microsecond latency
- 150 GB per second of throughput
- 10 million IOPS
These results were achieved with standard flash memory NVMe SSDs. The numbers will get even better when Dell EMC adds Optane-based storage class memory (SCM) as a tier. Once SCM has been added to the array, Dell EMC’s fully automated storage tiering (FAST) technology will monitor array activity and automatically move the most active data to the SCM tier and less active data to the flash memory SSDs.
The intelligence of the tiering algorithms will be key to delivering great results in production environments. Indeed, Dell EMC states that, “Built-in machine learning is the only cost-effective way to leverage SCM”.
HPE “Memory-Driven Flash” uses Storage Class Memory as Cache
HPE is one of many vendors taking the caching path to integrating SCM into their products. It recently began shipping Optane-based read caching via 750 GB NVMe SCM Module add-in cards. In testing, HPE 3PAR 20850 arrays equipped with this “HPE Memory-Driven Flash” delivered:
- Sub-200 microseconds of latency for most IO
- Nearly 100% of IO in under 300 microseconds
- 75 GB per second of throughput
- 4 million IOPS
These results were achieved with standard 12 Gb SAS SSDs providing the bulk of the storage capacity. HPE Memory-Driven Flash is currently shipping for HPE 3PAR Storage, with availability on HPE Nimble Storage yet in 2019.
An advantage of caching approach is that even a relatively small amount of SCM can enable a storage system to deliver SCM performance by dynamically caching hot data, even when it is storing most of the data on much slower and less expensive media. As with tiering, the intelligence of the algorithms is key to delivering great results in production environments.
The performance HPE is achieving with SCM is good news for other arrays based on caching-oriented storage operating systems. In particular, ZFS-based products such as those offered by Tegile, iXsystems and OpenDrives, should see substantial performance gains when they switch to using SCM for the L2ARC read cache.
What is Best – Tier or Cache?
I favor the caching approach. Caching is more dynamic than tiering, responding to workloads immediately rather than waiting for a tiering algorithm to move active data to the fastest tier on some scheduled basis. A tiering-based system may completely miss out on the opportunity to accelerate some workloads. I also favor caching because I believe it will bring the benefits of SCM within reach of more organizations.
Whether using SCM as a capacity tier or as a cache, the intelligence of the algorithms that automate the placement of data is critical. Many storage vendors talk about using artificial intelligence and machine learning (AI/ML) in their storage systems. SCM provides a new, large, persistent, low-latency class of storage for AI/ML to work with in order to deliver more performance in less space and at a lower cost per unit of performance.
The right way to integrate NVMe and SCM into enterprise storage is to do so–as a tier, as a cache or as both tier and cache–and then use automated intelligent algorithms to make the most of the storage class memory that is available.
Prospective enterprise storage array purchasers should take a close look at how the systems use (or plan to use) storage class memory and how they use AI/ML to inform caching and/or storage tiering decisions to deliver cost-effective performance.