[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: scheduler policy to spread pods

Hi, thanks for that suggestion. I took a look, but tit seems it isn't quite what's needed.
It looks likes pod (anti)affinity is a binary thing. It works for the first pod on the node with/without the specified label, but it doesn't ensure an even spread when you schedule multiple pods.

In my case I scheduled pods using an antiaffinity preferredDuringSchedulingIgnoredDuringExecution rule applying across 3 nodes and that made sure that the first 3 pods went to separate nodes as expected, but after that the rule seemed to not be applied (there were no nodes that satisfied the rule, but as the rule was 'preferred' not 'required' the pod was scheduled without any further preference). So that by the time I had 6 pods running 3 other them were on one node, 2 on another and only 1 on the third.

So I suppose the anti-affinity rule is working as designed, but that its not designed to ensure an even spread when you have multiple pods on the nodes.

On 04/07/18 12:16, Joel Pearson wrote:
Here’s an OpenShift reference for the same thing.

On Wed, 4 Jul 2018 at 9:14 pm, Joel Pearson <japearson agiledigital com au> wrote:
You’re probably after pod anti-affinity? https://kubernetes.io/docs/concepts/configuration/assign-pod-node/#affinity-and-anti-affinity

That lets you tell the scheduler that the pods aren’t allowed to be on the same node for example.
On Wed, 4 Jul 2018 at 8:51 pm, Tim Dudgeon <tdudgeon ml gmail com> wrote:
I've got a process the fires up a number of pods (bare pods, not backed
by replication controller) to execute a computationally demanding job in
What I find is that the pods do not spread effectively across the
available nodes. In my case I have a node selector that restricts
execution to 3 nodes, and the pods run mostly on the first node, a few
run on the second node, and none run on the third node.

I know that I could specify cpu resource requests and limits to help
with this, but for other reasons I'm currently unable to do this.

It looks like this is controllable through the scheduler, but the
options for controlling this look pretty complex.
Could someone advise on how best to allow pods to spread evenly across
nodes rather than execute preferentially on one node?

users mailing list
users lists openshift redhat com

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]