hyper-parallel continuum architecture
flipping heck it's a bloody time machine!?
Infinidat is aiming to replace VMAX/DS8000/VSP-type arrays with its Infinibox product, and it’s making good progress, having shipped more than 200PB of installed capacity in two years. An Infinidat briefing at VMworld added more information to what we have written here, here and here, about founder Moshe Yanai’s third attempt …
As a CTO, that isn't a very smart thing to say. It is downright irresponsible to call out storage controller functionality running on the server as not a good datacenter practice. You are after all doing nothing more than what DDN can deliver -- large capacity/high sequential B/W SANs.
> CTO Brian Carmody said: “Enterprises want packaged systems. SW-defined storage is like communism. It looks really good on paper but not so good in practise. There’s integration risk for customers. It’s the most un-cloud-like proposition you can make.”
This looks awfully similar to how Nimble does their data organization on disk (reads cached on DRAM/Flash and data is checksummed, coalesced before written to disk with parity using a log structured model).
The differences seem superficial ... non-deterministic drive selection. How do do they do garbage collection to avoid a whole bunch of junkie data from causing massive fragmentation issues?
How are they ensuring consistency on the front end? Is it using scale-out structure similar to Nutanix (3 nodes and up)? It would make sense to have the storage controllers share metadata using an efficient datastore like Redis or Cassandra with added Paxos functionality (Nutanix guys documented this in the Nutanix bible).
The way data is stored is unique. Incoming data is split into 64KB sections and the system is, below the protocol drivers, simply a collection of 3.125 billion of these sections. Concepts such as hosts and volumes don’t apply to the caches and disk drives. Each section is given a checksum and an activity vector (AV) score.
This AV score relates to the IO activity for a section and a heuristic map of inferred relationship to other sections, with the aim being to keep an application’s hot working set data in cache as its workload develops. The aim is to have all read requests be served from DRAM or NAND cache.
The system’s software groups 14 sections with similar AV values into a stripe, and two parity sections are computed to create a 16-way stripe for persistence to disk.
Next, 16 disk drives are chosen from the disk pool and the 16-way stripe is written. We’re told there is non-deterministic drive selection, log-structured writes and statistically uniform utilisation of drives.
Nimble went a traditional RAID under-pinning. Even went to triple parity in 2.1. The RAID arrays in Infinibox are 14+2 64k chunks dispersed as described. Caching appears similar but there are a number of similar caching schemes with SSD at this point. Nimble has standby controllers. Seriously? That's kind of lame. Infinibox is geared towards the Enterprise with mainframe support coming soon. Yeah, many Fortune 500s still support mainframes and some have quite a few of them. Regarding fragmentation, I don't think they fear it but embrace it. Writes hit "idle" disks (idle being relative but with 480 disks, some must be more idle than others) and reads are mostly cache hits.
@Robisrob == you didn't get the point of the OP.
Nimble systems are using write coalescing on to a log structured file system (with N+2 or N+3 RAID. The LFS make it possible to get parity without the read-modify-write penalty). Infinibox is using an erasure coding style RAID (N+2) chunking using a similar scheme. You can forget the tie in to disk geometries as rebuilds are data centric in both cases.
Also don't confuse the front-end/back-end issues. I am talking about data organization. Mainframe support isn't rocket science. EMC/IBM/HP/HDS want you to think it is, but it's FC with lot more validation.
Read the article and elsewhere, Infinidat is doing log structured writes also.
And as they describe it , virtual RAID groups. Each disk participating in numerous RAID groups, more here:
And no I wasn't implying that mainframe support is rocket science. It's just that the two have different target audiences. Without mainframe support, is it Enterprise (yes Enterprises run a whole bunch of kit, maybe a fairer description would be it isn't targetted for the high-end Enterprise?) But yeah, Nimble is interesting. Do you think they'll ever turn that standby controller into an active controller so it too can participate in serving IO?
Biting the hand that feeds IT © 1998–2020