Setting Pod Resource Limits and Requirements

djedje

Hello,

All is fine but still one point is unclear for me.

1. step one

I set a container resource like it

 resources:
          limits:
            cpu: "2" # with a limit of 2 cpu
            memory: "2500Mi"
          requests: 
            cpu: "1" # the container request 1cpu
            memory: "1950Mi"
        args:
        - -cpus
        - "1" # we set 1 cpu to the container
        - -mem-total
        - "1950Mi"
        - -mem-alloc-size
        - "100Mi"
        - -mem-alloc-sleep
        - "1s"

At run time i watch the describe of the worker node

Namespace Name CPU Requests CPU Limits Memory Requests Memory Limits AGE
--------- ---- ------------ ---------- --------------- ------------- ---
default stressmeout-797fb5bc6b-zqdwg 1 (50%) 2 (100%) 1950Mi (26%) 2500Mi (33%) 28s
kube-system calico-node-hzm52 250m (12%) 0 (0%) 0 (0%) 0 (0%) 25d
kube-system kube-proxy-mvwvk 0 (0%) 0 (0%) 0 (0%) 0 (0%) 25d
Allocated resources:
(Total limits may be over 100 percent, i.e., overcommitted.)
Resource Requests Limits
-------- -------- ------
cpu 1250m (62%) 2 (100%)
memory 1950Mi (26%) 2500Mi (33%)

I can see that i use 50%/62% of the cpu, seems to make sens as there is 1 cpu requested and 1 set ('args:1') on a limit of 2 cpu
But when i run the top cmd on the worker node, i got 100% cpu used.

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
6913 root 20 0 2015532 1.918g 3116 S 99.7 26.3 7:01.63 stress

2.step 2

I drop down the cpu request of the container

resources:
          limits:
            cpu: "2" # with a limit of 2 cpu
            memory: "2500Mi"
          requests: 
            cpu: "0.5" # the container request only 500millicpu
            memory: "1950Mi"
        args:
        - -cpus
        - "1" # we attribute 1 cpu to the container

i watch the describe of the worker node

Namespace Name CPU Requests CPU Limits Memory Requests Memory Limits AGE
--------- ---- ------------ ---------- --------------- ------------- ---
default stressmeout-7569595655-wjd9h 500m (25%) 2 (100%) 1950Mi (26%) 2500Mi (33%) 2m56s
kube-system calico-node-hzm52 250m (12%) 0 (0%) 0 (0%) 0 (0%) 25d
kube-system kube-proxy-mvwvk 0 (0%) 0 (0%) 0 (0%) 0 (0%) 25d
Allocated resources:
(Total limits may be over 100 percent, i.e., overcommitted.)
Resource Requests Limits
-------- -------- ------
cpu 750m (37%) 2 (100%)
memory 1950Mi (26%) 2500Mi (33%)

ok, it's make sens the cpu is only 25% of the limit, and resource 750m

But here i don t understand:
i watch the process of the pod on the worker node running top, and it's still using 100% of the cpu... why ?

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
6913 root 20 0 2015532 1.918g 3116 S 99 26.3 7:01.63 stress

Thank you.

chrispokorni

Hi @djedje,

Exceeding resource limits does not guarantee a container's termination and/or a pod's eviction. Memory and CPU limits are treated differently. While consuming more memory that the Limit may trigger an OOMKill, consuming more CPU than the Limit will not trigger a kill of the container.

The documentation offers more details.

Regards,
-Chris

djedje

Hello @chrispokorni ,

Thank you for the help. The link is interesting but still my question remain.

A few other exemple

1-----------------

resources:
          limits:
            cpu: "1" 
            memory: "500Mi"
          requests: 
            cpu: "0.5" 
            memory: "200Mi"
        args:
        - -cpus
        - "0.5" 
        - -mem-total
        - "350Mi"

Here the pod does not even start, il report an Error status. Does the container need a minimum of 1 cpu tu run ?

2------------------

 limits:
            cpu: "1" 
            memory: "500Mi"
          requests: 
            cpu: "0.5" 
            memory: "200Mi"
        args:
        - -cpus
        - "1" 
        - -mem-total
        - "350Mi"

Here the container run as expected. But when i 'top' the worker node, the pod process use 100% of cpus

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
13313 root 20 0 324332 320464 3116 S 99.0 4.2 0:09.51 stress

3------------------

resources:
          limits:
            cpu: "2" 
            memory: "500Mi"
          requests: 
            cpu: "1" 
            memory: "200Mi"
        args:
        - -cpus
        - "1" 
        - -mem-total
        - "350Mi"

Instead of requesting 500mcpu, i request 1cpu (but the args: is still 1 as previous), the 'top' still report 100%cpus
I feel that strange as even i request 500mcpus or 1 cpu, the 'top' still monitor 100% of cpu consuming ???

4-------------------

resources:
          limits:
            cpu: "2" 
            memory: "500Mi"
          requests: 
            cpu: "1" 
            memory: "200Mi"
        args:
        - -cpus
        - "2" 
        - -mem-total
        - "350Mi"

I increase the attribution of cpus to 2, which now match the cpu's limit of the container, run 'top' and get 200% of consuming
cpus (the pod still run without issues).

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
21720 root 20 0 324332 320724 3120 S 188.4 4.2 0:42.43 stress

So in my case the minimum cpu the pod is using is 100% (what ever the cpu's requests) ...?

and under an 'args[-cpus]' of 1 the pod report an error status...?

Reading the link does not answer it. What am i missing ?

Thank you.

serewicz

Hello,

The limit exists to terminate a pod when it tries to use more resources than allowed. The request is how much to hold back from other pods such that it is always available to the pod. The pod could use more or less than the request, but not more than the limit.

Regards,

chrispokorni

Hi @djedje,

It seems that the stress image does not like decimal CPU arguments, such as "0.5" and "1.5", but works well with integers like "1" and "2".
But once the container is running, the kubectl top command returns the expected CPU and memory utilization for both the node and the pod.

Regards,
-Chris

djedje

Thank you.