Well, another Storage Field Day (#12) is in the books. Again, honored to be amongst the delegates, and along with a stellar crew including @SFoskett, @MrAntropologist and the brand new @KatKitzmiller, not to mention the ever professional @PrimeImage doing audio and video, we tackled a variety of solutions in the storage world. Such names as we’ve known for years, like Nimble and Intel as well as others we’d never known known like Excelero and Ryussi.
The tech was impressive, and we’ve seen solutions we’d never anticipated previously. A few of these simply blew my mind away.
I will start with Excelero. Imagine, if you will, that you had the ability to create a storage node or cluster consisting of Solid State disc. For example a storage node, in this example, consisting of 16 NVME discs. The NVME format is well suited for this type of work, because it sits directly on the PCI bus, thus much closer to the processor core than say, SAS or SATA which require some controller overhead for their connection. A total of sixteen discs each delivering roughly 281,000 iops, together delivered the full 4.5million iops. That actually means that the storage infrastructure itself, including the iSCSI stack did not add any overhead whatsoever to the disc.
Again, the storage node itself had 24bays, two sockets of Xeon with ample threading to support the purpose, but not at all expensive. Only 16 of these bays were filled, so this one device were it to be fully populated, could theoretically deliver 6.75Million Iops. All for roughly $15,000. The excitement over a product like this truly lit a fire in the eyes of the delegates attending. @RayLucchesi particularly could not contain his excitement, and that lasted for him days.
So, a bit about the technology: The disc was accessed using a version of RDMA, that Excelero created called RDDA, or NVME over Fabric. The iSCSI connectivity to the disc, also was a freshly written stack used by them, and it effectively caused no overhead to the disc, even when connecting over the 100Gb Ethernet. The networking itself, was commodity equipment as well, in this case Mellanox and is quite effective. The large amount of bandwidth provides for a Mesh of NVME
Data Protection and redundancy in the configuration works in “Rain” fashion, which essentially means that it’s provided across storage nodes, and even across racks. This is native to the configuration.
There is a “Topology Manager” element, which performs cluster management, ensuring HA, it manages the volume life-cycle, and failure recovery operations, and also manages the RAFT protocol which ensures consistency across disc, and nodes, and prevents “Split Brain.”
Scalability of this architecture is essentially without barrier.
As of yet, though we were led to believe that may be roadmapped, was Compression and Deduplication. Personally, I doubt that I’d be willing to sacrifice the performance necessary for these functions, in order to have Compression or Dedupe, but I’d be surprised if the fine folks at Excelero would write these additional functions into the code and allow them to sacrifice much in performance.
The storage overhead in the case of this distributed converged platform is so low, that the servers themselves, some which house the disc, or others which can be attached have still so much bandwidth that they could act as applications servers, as well as storage devices on the same nodes.
Astounding technology, and we were proud to be in attendance as the organization provided this overview to us and to the world as they came out of Stealth on the product. I cannot wait to see how something like this shakes up the industry.
The previously mentioned Ray Lucchesi beat me to the punch on a posting for this tech:
And, there are a host of great blog posts from my fellow #SFD12 delegates here at http://bit.ly/2ngDohq