Adding my 2cents. Inline with your questions.
-----Original Message-----
From: Axel Baudot <axel.bau...@protonmail.com.INVALID>
Sent: Friday, April 21, 2023 1:14 AM
To: users <users@cloudstack.apache.org>
Subject: Arguing for Cloudstack for a HPC oriented datacenter
Hello dear CloudStack users,
As a team sitting at a crossroad to chose a solution to provision ressources
for a HPC oriented datacenter: can a strong case be made for the use of
Cloudstack over Openstack, OpenNebula or Kubernetes?
In particular I have the following concerns :
1. The company provides infrastructure to clients, as well as full blown
project development, so it trying to set up both a public cloud with resource
billing and an internal development platform. This is the use case covered by
CloudStack, right? Or could the flexibility of OpenStack be needed?
>>Cloudstack is absolutely fit for purpose there, I've used it myself for
>>exactly that, HPC and HPTC.
2. It is also heavily geared toward HPC. I have seen little mention of HPC with
CloudStack, while it is heavily advertised in OpenStack world, due to the ties
with CERN and such. Can you think of limitations of CloudStack for this use
case? I am also interested in case studies or any reading material on this
combination.
>HPC has two sides:
1) a collapsed core backbone network, 100Gb+ interfaces, extremely low latency
switches, NVMe based storage arrays with Fibre Channel connectivity,
etc...etc... CloudStack is completely neutral to that and adds no performance
impact in these domains.
2) GPU processing - CloudStack has support for GPU passthrough and new GPU
models can be added relatively easily.
Distributed network processing - CloudStack now supports Tungsten Fabric,
so you can have routed traffic in extreme scales without bottlenecks. Or,
simply use pure L2 networks.
Specialized templates - CloudStack doesn't limit that
High CPU count - CloudStack doesn't limit that, it's only a hypervisor
limit
3. The team could standardize on containers soon (especially for dev
environments). I would tend to do a simple VM + Docker workflow. Is it standard?
>Absolutely standard
4. I am all for simplicity of operation and maintenance. I think CloudStack
could really shine here, right? I also think that Kubernetes should be avoided
because of this. Thing is, Kubeflow is getting some attention and we might have
to support it... Is CKS considered mature for production and a viable solution
in this case?
Thanks for reading, any food for thoughts will be very much appreciated.
Best,
Axel
