This post has been updated with details from HPE Discover 2018, where the use of Intel Optane is being described as HPE Memory-Driven Flash for existing 3PAR and Nimble products (see HPE blog).
There’s been a lot of coverage and potential hype around the impending release of Intel and Micron’s 3D-XPoint persistent storage technology. Now it seems we have a good example of how Optane (as Intel are branding it) could be deployed into existing storage platforms. During HPE Discover London (December 2016), the 3PAR team ran a demo showing the results of using an Intel Optane NVMe adaptor card as a caching layer for an all-flash 3PAR system. Optane is in a group of products being called “Storage Class Memory” because the addressing mode is at the byte rather than block level.
Initially Intel and Micron claimed 3D-XPoint to be 1000 times faster than NAND and have 1000 times higher endurance. Those bold statements are yet to be proven and in fact the performance figure can be debunked somewhat (link), but the technology will be faster than the NAND flash being used today. As a result, there’s an opportunity to use Optane as a caching layer that sits between DRAM and flash in order to improve I/O performance. That’s exactly what the 3PAR team have done.
3PAR already has a feature that can make effective use of flash in hybrid platforms. It’s called Adaptive Flash Cache and allows flash media to act as a “level-2” read cache to supplement DRAM caching when the active working set size of data is greater than the DRAM capacity. You can find a white paper on the details here (link). In all-flash systems, there’s no benefit having Adaptive Flash Cache as the read request can be taken directly from persistent data on flash. However with a medium faster than flash, the idea of adaptive caching becomes interesting again. Eduardo Duarte (HPE 3PAR Product Manager), chatted about the performance improvements that could be achieved with the use of Optane. Indicative figures show speed improvements that reduce typical I/O latency from 350μs to 200μs. Now there’s a caveat here. The storage fabric will introduce an impact to performance at this level of latency. Eduardo indicated that best results were seen with 16Gb/s Fibre Channel compared to 8Gb/s hardware, which isn’t surprising, bearing in mind the increase in throughput between these generations.
Update December 2018: The performance figures quoted above have been confirmed as the expected performance levels for 3PAR with SCM.
Sweating the Asset
The use of Optane in 3PAR is currently only a tech preview, but hopefully the technology will be available soon. There will be issues to resolve around power draw and how many devices can fit into a single chassis, so I expect Optane will initially be offered in the higher end platforms. What this announcement shows is another incremental improvement in sweating the 3PAR asset. It’s sometimes difficult to see what extra features could be added, however the 3PAR team manage to keep coming up with ideas.
The Architect’s View
The interesting thing about continuous incremental improvement is the ability to continues to retain existing customers and attract new ones. Contrast this to Dell-EMC’s DSSD, which is rumoured to have only around $6m in revenue from about a dozen sales (link). Admittedly EMC is attempting to create a new niche, however technologies like Optane may leapfrog the need for direct attach high performance storage. Instead this may be directly integrated into the server. Continuous improvement may not be as sexy as creating a new category, but will bring revenue earlier – 3PAR still appears to have a few years of life yet!
- Storage Field Day 11 Previews: Intel, HGST
- Storage Field Day 9 Preview: Intel
- Why 3D-XPoint SSDs Will Be Slow (TheSSDGuy.com, published 16 June 2016, retrieved 12 December 2016)
- 3PAR and Storage Class Memory: From Ludicrous Speed to Plaid (Around The Storage Block blog, published 22 September 2016, retrieved 12 December 2016)
- HPE 3PAR Adaptive Flash Cache (PDF, retrieved 12 December 2016)
- DSSD President quits Dell-EMC (The Register, published 9 December 2016, retrieved 12 December 2016)
Comments are always welcome; please read our Comments Policy first. If you have any related links of interest, please feel free to add them as a comment for consideration.
Copyright (c) 2007-2020 – Post #FC0F – Brookend Ltd, first published on https://www.architecting.it/blog, do not reproduce without permission.