Scaling of apps

[bar1]

Hi, so currently i have 2 scale servers running.
how do i use Kubernetes to scale apps between the 2?
do i need to replicate ix-applications between the 2 servers and set Desired Replicas to 2?
thanks

 

[sretalla]

how do i use Kubernetes to scale apps between the 2?

You don't (yet). That's coming in future releases.

 

[bar1]

You don't (yet). That's coming in future releases.

ahhhh ok can't wait!!!!

 

[bar1]

still not available?

 

[gsrcrxsi]

I have a sort of related question about the scaling, when it is eventually available.

will the scaling be additive? or redundant?

for example, say you have two identical systems each with 4 CPU cores. will a single app be able to utilize all resources from both systems (8 cores)? or will this just be for failover protection allowing one app to seamlessly exist on either system if one goes down?

just want to understand the limitations.

 

[truecharts]

I have a sort of related question about the scaling, when it is eventually available.

will the scaling be additive? or redundant?

for example, say you have two identical systems each with 4 CPU cores. will a single app be able to utilize all resources from both systems (8 cores)? or will this just be for failover protection allowing one app to seamlessly exist on either system if one goes down?

just want to understand the limitations.

There is no App clustering in any way, shape or from in the current release of SCALE.
Nor is there a completed formal design draft internally at iX Systems.

But generally speaking about kubernetes it's neighter.
It's just spreading containers (of which there can be multiple, including copies in an App) over the hosts.

 

[gsrcrxsi]

Thanks. That’s what I was getting at. If it would be (I’m aware there’s no clustering NOW) the ability for a single app to use all resources of the entire cluster.

Based on your reply, sounds like it won’t be possible.

 

[sretalla]

sounds like it won’t be possible.

Not currently...

If the "app" is something that can run multiple instances of a container (kubernetes pod scaled to greater than 1 instance across multiple hosts), that will be possible in future releases.

 

[gsrcrxsi]

Yeah sounds like previous implementations

This is a little off topic, but outside of TrueNAS, is there any OS or management software that DOES allow clustering of multiple hosts where the resources are all available to a single app? Where two 4-core hosts could pool together and some application could have all 8 cores to use for computation tasks.

 

[truecharts]

Yeah sounds like previous implementations

There never has been application clustering on SCALE, not now and not previously.

Yeah sounds like previous implementations

This is a little off topic, but outside of TrueNAS, is there any OS or management software that DOES allow clustering of multiple hosts where the resources are all available to a single app? Where two 4-core hosts could pool together and some application could have all 8 cores to use for computation tasks.

As explained by @sretalla it is theoretically possible with kubernetes IF the application/container(s) you deploy support it.

 

[sretalla]

What you're talking about might be closer to grid computing... a very specific setup which allows multiple hardware to act on a single set of data with all CPU and memory pooled, but you can't just run any old application on a grid, it needs to have been written for it.

Examples of Grid Computing - Real World • ThePicky

Grid computing is making big contributions to scientific research, helping scientists around the world to analyze and store massive amounts of data. SETI

www.thepicky.com

What is Grid Computing? | Componenets, working of Grid Component

This is a guide to What is Grid Computing? Here we discuss how Grid Computing works, components, and use case in detail information.

www.educba.com

No "legacy" application (one that wasn't written to be cloud or cluster native) will be able to run across any kind of cluster which involves separate nodes. At best, you would be able to have higher availability by failing the application from one node to another, but without the right application structure, it won't cope with more than one host.

 

[bar1]

so for my home lab, i just need something if a server is down even if it's not proper scaling/clustering.
I assume i can replicate ix-applications between my 2 truenas servers?
is it possible to replicate only SOME apps?

 

[sretalla]

If you want kubernetes clustering, you can start with the new bluefin nightly images (and learn the CLI for that as the TrueCOmmand UI isn't yet available for it).

As we already said, clustering isn't yet available on the released version.

You can replicate datasets, but the datasets for applications is a bit complicated, so you'll need to do some research to understand how to make that work for you. (or wait for the release of Bluefin)

 

[gsrcrxsi]

There never has been application clustering on SCALE, not now and not previously.

I'm aware. I mean previous implementations of kubernetes that I've read about (RPi clusters), not TrueNAS specifically.

 

[sretalla]

You can always build a cluster of docker VMs using something like docker swarm or Rancher (or if you want to pay, VMware Tanzu).

Those options don't really play directly on SCALE as there's iX's own front-end doing that job that doesn't yet offer the options you're after, but already consuming the kubernetes endpoint.

 

[gsrcrxsi]

thanks, I'm not really looking for anything right now, just find the concepts interesting if I ever wanted it in the future.

a potential application for something like this that I had in mind was BOINC. BOINC performance scales linearly with number of threads since each job only occupies a single thread. if you could pool together two systems for greater number of threads you can run more jobs concurrently. At the moment you'd need to install separate instances for each host. there are people with hundreds or thousands of hosts and indeed they have a separate application instance for each one. deployment and management is a large undertaking. since it doesnt require all threads to work on the same job, node to node latency should be no issue. think of something like a middle man that is between the project and the various hosts,it gets work and distributes it to any available thread no matter what node it's on. it just sends the job to an available core, then the results come back to be sent back to the project. to the project it looks like a single system when it's really many on the backend. currently the BOINC application itself can only really detect and use the resources of the host it's on. but if you could group multiple systems together it might be an interesting solution for mass deployments.

I do a lot of BOINC processing, but honestly my focus is on GPU computing, though I do a bit of CPU crunching as well with a dual 48-core (192 thread) EPYC system. I only have a handful of hosts at the moment so it's not too troublesome for me to just have separate instances. I was just thinking out loud about what is currently possible.

 

[truecharts]

thanks, I'm not really looking for anything right now, just find the concepts interesting if I ever wanted it in the future.

a potential application for something like this that I had in mind was BOINC. BOINC performance scales linearly with number of threads since each job only occupies a single thread. if you could pool together two systems for greater number of threads you can run more jobs concurrently. At the moment you'd need to install separate instances for each host. there are people with hundreds or thousands of hosts and indeed they have a separate application instance for each one. deployment and management is a large undertaking. since it doesnt require all threads to work on the same job, node to node latency should be no issue. think of something like a middle man that is between the project and the various hosts,it gets work and distributes it to any available thread no matter what node it's on. it just sends the job to an available core, then the results come back to be sent back to the project. to the project it looks like a single system when it's really many on the backend. currently the BOINC application itself can only really detect and use the resources of the host it's on. but if you could group multiple systems together it might be an interesting solution for mass deployments.

I do a lot of BOINC processing, but honestly my focus is on GPU computing, though I do a bit of CPU crunching as well with a dual 48-core (192 thread) EPYC system. I only have a handful of hosts at the moment so it's not too troublesome for me to just have separate instances. I was just thinking out loud about what is currently possible.

For BOINC there are great kubernetes solutions possible, but that does not require what you are suggested you needed previously.
Please just wait till people that know how to do the things you want, offer an App that does what you want it to do.

Anyway, that is of totally irrelevant at the moment as it's not something that is supported anyway.

 

[truecharts]

If you want kubernetes clustering, you can start with the new bluefin nightly images (and learn the CLI for that as the TrueCOmmand UI isn't yet available for it).

As we already said, clustering isn't yet available on the released version.

You can replicate datasets, but the datasets for applications is a bit complicated, so you'll need to do some research to understand how to make that work for you. (or wait for the release of Bluefin)

There is no kubernetes clustering in the Nightly either.
The internal design draft was not even finished as-of 2 days ago.

Nor is it 100% clear what is going to be TrueCommand based and what's not.

 

[sretalla]

There is no kubernetes clustering in the Nightly either.

I was thinking about this post: https://www.truenas.com/community/threads/truenas-scale-bluefin.99254/post-684750 when I wrote that...

 

[truecharts]

I was thinking about this post: https://www.truenas.com/community/threads/truenas-scale-bluefin.99254/post-684750 when I wrote that...

Mistake by @Kris Moore it seems, the post he linked was about storage clustering, which is also part of TrueCommand already.
(check the dates).