Cluster hardware

[bbuster]

I have a server (HPE Proliant ML350 Gen 9) with 2x E5-2667 v3 and 128GB ram.
I want to build a cluster and see a HPE Proliant ML350 Gen 9 with 2x E5-2687W v3 and 256GB ram


Can this be clustered together (i know i still need a 3d one).
My current server has 2x Samsung PM893 2.5" 1920 GB SATA III V-NAND TLC

The server i can buy has:
4x HP 2.4TB SAS 12G 10K HDD
2x HP 800GB SAS 12G SSD
2x HP 480GB SATA 6G SSD

Can i use these HDDs or better buy the same Samsung PM893?


And can i cluster it later when everything is running already on my server or does it need to be setup as a cluster before creating virtual machines / dockers?

 

[UdoB]

Can this be clustered together

Sure!

And can i cluster it later when everything is running already on my server or does it need to be setup as a cluster before creating virtual machines / dockers?

The new to-be-added node shall have no VMs/LXc on it - it needs to be "empty" for a clean join.

Also https://pve.proxmox.com/pve-docs/pve-admin-guide.html#chapter_pvecm is for you

 

[UdoB]

Can i use these HDDs or better buy the same Samsung PM893?

Yes. You can use HDDs. AND also: Yes, PM893 is a better choice. It always depends on your need/expectations and money availability.

Note that you can start with HDDs and replace them later by physically compatible SSDs...

My personal rule of thumb - without knowing your use case - is: if you want a reasonable speed build a ZPool with multiple striped mirrors (aka Raid10) and add one pair of "Enterprise"- SSD (also as a mirror) as a "Special Device". Do NOT do a "Cache" or an "SLog".

 

[bbuster]

Thank you for your reply, i will read your documentation but still a few questions.

I now have 1 server with a few virtual hosts. I can use this as a "main" machine later and add 2 other empty machines to create the cluster without removing the hosts on the main machine?
And i don't understand the "dedicated NIC" part. Can i find any additional information about this anywhere?
The cluster needs to be installed on the same location right?

The PM893 are sold to me as "enterprise" ssd's, is this what you refer to or do i need better ssd's?
I would like to prevent downtime because it will be used for multiple production web applications

 

[UdoB]

I now have 1 server with a few virtual hosts. I can use this as a "main" machine later and add 2 other empty machines to create the cluster without removing the hosts on the main machine?

Yes.

First you formally create a cluster on a single machine - that one containing VMs already. Then you can add additional (empty) nodes to this cluster.

And i don't understand the "dedicated NIC" part.

Reference? Networking is a large topic. One usual recommendation is to have a physically separate network for corosync - so it can not be congested by heavy traffic on the "other" network - and also adding redundancy for this. And one network for Storage and one ...

While you can run a cluster on one single network it is definitely better to separate specific traffic.

(And then there is still no redundancy with separate NICs and separate physical Switches for one logical network.)

The cluster needs to be installed on the same location right?

The network connectivity between the nodes needs to have low latency. If you can connect two cities with a "ping time" of 2ms it would be fine. If you are at home with DSL and get response times of -lets say- 20ms it won't work reliably.

The PM893 are sold to me as "enterprise" ssd's, is this what you refer to or do i need better ssd's?

It is fine! The datasheet says "Samsung’s PM893 and PM897 SSDs include enterprise power loss protection (PLP" and "Endurance: 1.3 DWPD for 3 years"

 

[esi_y]

@UdoB Do you mind me asking (emphasis mine):

if you want a reasonable speed build a ZPool with multiple striped mirrors (aka Raid10) and add one pair of "Enterprise"- SSD (also as a mirror) as a "Special Device". Do NOT do a "Cache" or an "SLog".

Why? (Especially the cache.)

 

[UdoB]

Why? (Especially the cache.)

Because I labelled it my personal rule of thumb

First maxout RAM. Put in as much as technically possible. Then use arc_summary.pl and watch the hit/miss ratio. Then ask a person with more knowledge than I have. Then decide if you need an additional cache.

A cache device needs Ram to work, so the ARC will shrink. There might be configurations out there which got slower after adding a cache...

Disclaimer: not tested and verified by me...

 

[esi_y]

Because I labelled it my personal rule of thumb

I see.

First maxout RAM. Put in as much as technically possible. Then use arc_summary.pl and watch the hit/miss ratio. Then ask a person with more knowledge than I have. Then decide if you need an additional cache.

This is true on any system, but it can be cost prohibitive.

A cache device needs Ram to work, so the ARC will shrink. There might be configurations out there which got slower after adding a cache...

This is a good point, probably not a good idea to add zcache to system chronically already low on RAM indeed.

Disclaimer: not tested and verified by me...

Thanks for the reply!