Reply to post: Centralized all flash is just an impressively bad idea.

Enterprise storage sitrep: The external array party is over


Centralized all flash is just an impressively bad idea.

Ok... so, you have SSD media capable of 2GB/sec read and write. Assuming almost zero overhead, that means to transport that 2GB/sec would require 20Gb/sec (as NRZ coding and clock recovery will always consume the difference). This means that over a single 100Gb/sec network connection, you can achieve maximum performance of only 5 storage devices.

Ok, so you're using NVMe over a PCIe switched fabric. Even if this were the case, you're probably still maxing out at 20 storage devices. We're not even considering processing overhead. So, let's assume you can read and write an aggregate of 40GB per second across your array. You would need a storage device able to handle compression, deduplication, error correction, possibly erasure coding etc... for 40GB per second. Can it be done... in ASICs sure!!! no worries! Of course ASICs can't be upgraded so unless you're 10000% sure that the array manufacturer uses nothing be absolutely perfect engineers who never make mistakes, you'll need something probably software based.

No... Cisco, HP, Huawei, EMC, NetApp... none of these guys can deliver anything that can possibly make good use of all flash anything. Not only that, but even using awesome tech like XPoint, the latencies required to handle processing centralized storage would be far too long to waste time on this project.

Buy local consumer grade M.2 storage (or XPoint if you're truly wasteful) and put four of them in each machine. Then add some spinning disk for capacity to each server. Then run a proper share-nothing storage system. If you're really bad at your job, you can use systems that manage virtual disks and block based storage. This is how "Storage Experts" do it.

Or if you're really good at your job, build your storage infrastructure around your systems and you can actually make do with far less overhead.

Quit throwing away all your money on things like storage arrays and start designing your storage properly. This means databases, object storage, etc... it's 2018. Any project you start now that you put even the slightest effort into will run until 2020 at least. Might as well do it right.... or you throw good money after bad and be "a storage and VM expert"

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon


Biting the hand that feeds IT © 1998–2020