Questions about SSD NVMe for 1C?

Question

System administration

Good afternoon!
Comrades, please share your experience / knowledge about using SSD NVMe for 1C.
1. How to make RAID 1 on NVMe SSD in ESXi?
Thoughts:
a) It turns out there is no point in using Raid controllers in this case. in the case of NVMe, we are just trying to get rid of intermediaries with wires that slow down the speed, switching to PCIe with direct access to the processor.
b) It remains only to forward 2 NVMe SSDs to the VM and create a software raid already inside? (for processors that support PCI-passthrough)
c) But a simple Raid controller will only need Raid 1 for ESXi itself and a VM without any special disk loads. Without a Raid controller, it's impossible to make Raid 1 for ESXi, right?
2. For NVMe SSDs, only raid 1 is needed for reliability, and the rest of the raids are not economically justified?
Thoughts:
a) For productive work with server 1C for many users, first of all, a random reading speed of a large number of small blocks is needed, and secondly, a random writing speed of a large number of small blocks. Or vice versa? At peak load - what is usually the bottleneck first?
b) All levels of raids give only an advantage in the linear read / write speed, which in our case is practically not needed. Therefore, a raid is needed only for reliability (mirror) or an increase in the array (although what prevents you from doing a lot of Raid1 if you don’t directly need a single large space?)
3. Because Enterprise level NVMe SSDs are quite expensive, is it possible to somehow save money, but still have reliability?
For example, somehow use slower SSDs, but many times cheaper.
Raid 1 is no longer needed, because slower SSDs will be a bottleneck.
Is there any solution when the NVMe SSD is the first drive and always gives maximum performance (both during normal operation and with short-term peak loads), while other SSDs are overwritten with a queue in only peak loads? (the rest of the time they will work synchronously, because the peak load occurs only 5% of the time of the entire work of the disks.).
If the NVMe SSD dies, then the current data queue is written from the buffer (some RAM disk is acceptable) to another SSD and the system continues to work on an SSD with lower performance until a new NVMe SSD is inserted.
---
Not very technically described, but something like this. At the same time, not 2 expensive disks will wear out, but only 1, the second disk in stock lies for a quick replacement. Let less expensive SSDs wear out, change, they are still many times cheaper.
Or there is no point in this, because. Hi Enterprise SSD TBW is outrageous and we will replace a bunch of cheaper SSDs over its lifetime, what will the cost be the same in the end?
4. How to hotswap an NVMe SSD?
Intel has U.2. SuperMicro also has solutions in the box.
Are there any adapters that can be hot-swapped, but at the same time without loss of performance due to wires / adapter?
5. Will an NVMe SSD work on an old X9DRi-LN4F+ https://www.supermicro.com/products/motherboard/xe... ?
Maybe someone has experience?

Answer 1

Raid controllers make no sense to use in this case

Why not?
What are the wires. Just plug the controller into PCI - disks into the M2 slots of the controller and that's it.
Maybe a soft raid.
In general, in the case of a client-server scenario, the DBMS effectively caches data in memory, which reduces random disk load. But it depends on the amount of RAM, the size of the database, the number of clients.
Without knowing this, it is difficult to predict the load.
But in general, a well-tuned SQL is characterized by a load close to linear.
Regarding the raid - yes, it makes sense only in a mirror raid to ensure fault tolerance.
Don't worry too much about disk wear - there are not such volumes of recording as a rule.
Look towards 3DXPoint memory - Intel optan and others.
There are no problems either with the resource, or with TRIM and other NAND memory troubles.
Before installing expensive disks, analyze the current load, maybe the HDD can handle it there.
Regarding enterprise disks and the usual differences in the following -
1) support for different server modes.
2) an increased amount of reserve (no one bothers you to do this on a regular disk)
3) more memory and a more powerful processor - where the user will be bent under load, the enterprise will plow.
And 1s servers are different - somewhere the load is almost zero, somewhere high. You need to know the number of users, the size of the database, whether users are constantly actively working with the database or not.

Answer 2

But a simple Raid controller will only need Raid 1 for ESXi itself

ESXi does not work with softtrades. She just doesn't see them . She sees the discs that make them up.

Answer 3

More operatives, sql loves it.
Disks are regular ssd sata. Change as you die... what will happen to me in 3 years, not earlier... (of course, it depends on the amount of "changeable" information.
If for ESXI - I would do this - Forward disks directly to the virtual machine, assemble a soft raid there ..
Or KVM virtualization on linux... Again, there is a soft raid by the system itself

Answer 4

Colleagues, We have a working prototype of a 1C server on AMD + NVMe, we can give it for testing. Works very fast!!!