Is SCALE a good idea for compute node?

[Tenek]

HI folks,

I need to build a fast node for algorithmic computations. I'm thinking to use 2x Xeon GOLD 6136 for that.
Currently I'm using Rancher and all my deployments and workflows are kubectl based. My rancher installation is based on ubuntu nodes, which are VMs inside TrueNas Core, so I pretty slow with all these overheads.

I think I have two options:
- Use SCALE on the compute node and jus deploy my services there? I guess I can fine tube how many cores/threads can be used by pods and pods will have access to it without any overheads?
- Install ubuntu on the compute node and add it as a node into my current rancher orchestration?

Also, with SCALE:
- can I assign DIRECT access to NVME disks for the certain nodes?
- can I run VMs if necessary (similar to VMs in CORE)?

And finally, for higher performance and future proof should I go, for example, with 2x Skylake Xeon Gold 6136 or EPYC 7401P, or save some money and stick with two Xeon E5 2(6|7|8|9)xx?

Thanks!

 

[truecharts]

To be clear: SCALE itself CANNOT as a kubernetes cluster, only as a self-contained single-node cluster.
It can also not join an existing cluster.

VM's are supported though...

 

[Arwen]

@Tenek - While their is great promise for TrueNAS SCALE, it's barely out of Beta testing. And does not support cluster, yet.

It can be frustrating that their are basically no other options, for clustered Kubernetes with reliable storage. But, that was the reason why iXSystems started TrueNAS SCALE. It just will take them time to get the reliability up to TrueNAS Core standards. And then add the various cluster features.

 

[Tenek]

Thanks for the feedback, folks. I know that SCALE is its early stages. Single self-contained cluster/node might be fine for me. I like an idea to have less overhead, i.e. SCALE->PODS is better than CORE->virtual VMs->PODS. I'm just not sure how disk/CPUs resources are managed, as I mentioned ideally I would like to have direct access to it from certain workloads. I guess I need to try and play with it.