[
https://issues.apache.org/jira/browse/YUNIKORN-1120?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Craig Condit updated YUNIKORN-1120:
-----------------------------------
Description:
Currently, only absolute values for resources are allowed in configurations
read by the core. We should allow unit specifiers as well (M, Mi, G, Gi, etc.).
>From YUNIKORN-165:
Internally we must move to integers for the representation, just to fix the
issue described in YUNIKORN-1105. We should be able to handle about 9 exa byte
as the maximum size of a cluster using a standard int64 or 18 exa byte if we
use an unsigned integer. Not sure if we will overflow that in any resources
quickly.
What I meant with the staying closer to what K8s does is to allow specifying
the resources with a quantity suffix specifically in the configuration. This is
what we currently require:
{code:java}
resources:
max:
vcore: 21000
memory: 610000 {code}
That specifies 21 cpu's (base is milli vcores) and 610GB for memory (base is
1MB). Not that simple to interpret. We should allow 610G, and like with the
shim side we need to convert that to B. Same for the vcore, 21 should be
converted to 21000 m cores etc.
I would like to see us support a configuration like this:
{code:java}
resources:
guaranteed:
vcore: 2500m
memory: 600G
max:
vcore: 10
memory: 1T{code}
This is independent of how we represent the data in the REST specification.
REST should align with the base suffix: milli for vcores (base == 1e-3) and
bytes / undefined for others (base == 1). So never a quantity suffix in REST,
vcores always as milli vcores and all other resources as bytes or undefined.
was:Currently, only absolute values for resources are allowed in
configurations read by the core. We should allow unit specifiers as well (M,
Mi, G, Gi, etc.)
> Allow unit specifiers in core configs
> -------------------------------------
>
> Key: YUNIKORN-1120
> URL: https://issues.apache.org/jira/browse/YUNIKORN-1120
> Project: Apache YuniKorn
> Issue Type: Improvement
> Components: core - common, core - scheduler
> Reporter: Craig Condit
> Assignee: Craig Condit
> Priority: Major
>
> Currently, only absolute values for resources are allowed in configurations
> read by the core. We should allow unit specifiers as well (M, Mi, G, Gi,
> etc.).
> From YUNIKORN-165:
> Internally we must move to integers for the representation, just to fix the
> issue described in YUNIKORN-1105. We should be able to handle about 9 exa
> byte as the maximum size of a cluster using a standard int64 or 18 exa byte
> if we use an unsigned integer. Not sure if we will overflow that in any
> resources quickly.
> What I meant with the staying closer to what K8s does is to allow specifying
> the resources with a quantity suffix specifically in the configuration. This
> is what we currently require:
> {code:java}
> resources:
> max:
> vcore: 21000
> memory: 610000 {code}
> That specifies 21 cpu's (base is milli vcores) and 610GB for memory (base is
> 1MB). Not that simple to interpret. We should allow 610G, and like with the
> shim side we need to convert that to B. Same for the vcore, 21 should be
> converted to 21000 m cores etc.
> I would like to see us support a configuration like this:
> {code:java}
> resources:
> guaranteed:
> vcore: 2500m
> memory: 600G
> max:
> vcore: 10
> memory: 1T{code}
> This is independent of how we represent the data in the REST specification.
> REST should align with the base suffix: milli for vcores (base == 1e-3) and
> bytes / undefined for others (base == 1). So never a quantity suffix in REST,
> vcores always as milli vcores and all other resources as bytes or undefined.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
