Help with vm Win srv 2019 too slow

[Letizia]

Hi friends, I did a lot of tests and bought an enterprise disk to do the tests you suggested.
In the photo with two benchmarks you can see:
on the left side the Samsung ssd 870 EVO disk and
on the right the new Samsung PM1643A 2.5" 960 GB SAS disk.

In single photo (with only one benchmark) we can see Samsung 870 EVO disk directly to p440 controller of HP380 Gen9 - NO Proxmox installed, just Hardware and Windows Server Directly installed from ISO image.
It seems I'm doing something wrong, or is the problem the disks together with the virtualizer/Ceph?
If I try another hypervisor could I have the same problem?

I share every test with the hope of being of help to other people, I hope I don't post too many unhelpful posts.

Thank you for your interest and any ideas!

 

[Neobin]

Have a read here: [1] for why you want SSDs with PLP for things/workloads, that (at least partly) (heavily) do / depend on sync writes, like for example: Ceph, ZFS and databases.

Who does not do sync writes? Right, CrystalDiskMark.
But I already mentioned before, that this is not an appropriate benchmarking tool (at least in this case)...

PS.: And of course, a single enterprise SSD with PLP in a Ceph (or ZFS) pool with the remaining SSDs being still consumer SSDs without PLP, will bring you basically nothing... (The chain is only as strong as its weakest link. )

[1] https://forum.proxmox.com/threads/r...ritical-mini-server-system.120097/post-543371

 

[Letizia]

Hi everyone,
I hope this, in addition to being an update, can help those who have had the same problem as me.

I did some more testing in my machines and bought new boards.
After reading your suggestions, my idea was to put a PCIe X16 card with 4 cables and 4 NVME u.2 hard drives.
Unfortunately, the card (img 01) doesn't seem to be recognized by the DL380 G9 system (I still don't understand if the problem is the card or the cables! the disks remain off).
Then I bought 1 PCIe card with integrated dual u.2 sockets, but I could only see one disk (still the same NVME u.2 disk) .
The last test was to insert 3 PCIe cards with single u.2 socket with the relative disks.
The disks used for all tests were always and only Kingston DC1500M.

I am attaching images of the components and results visible inside proxmox.
I plan to buy another 6 boards and another 6 hard disks to reach 9 disks in total on 3 nodes (3+3+3).
I did some tests (no benchmark) and I noticed a big improvement with VMs.

In the fourth image we can see the difference between the DC1500M disks with single slot PCIe card and the previous samsung SSDs.

Any suggestion is welcome.

Thanks!
Letizia

 

[Neobin]

Then I bought 1 PCIe card with integrated dual u.2 sockets, but I could only see one disk

Most likely because the mainboard (and its bios/UEFI) needs to support PCIe bifurcation and if it does, you need to configure the corresponding PCIe slot(s) accordingly (in the bios/UEFI).

I strongly assume, this is also true for that 4-port adapter card, but without (configured) PCIe bifurcation, it should leastways also show (only) one SSD, I would guess.

 

[Letizia]

Thanks @Neobin,
I've looked everywhere in the BIOS, but haven't found any about PCIe bifurcation settings.
Yes, you are right, in fact only one disk was visible.

Today i am adding 3 PCIe u.2 adapter cards for each node, I will remove the 3 old SSD drives and I hope it solve my problem with speed.
In the image you can see my current situation.


Letizia