[
https://issues.apache.org/jira/browse/YUNIKORN-2325?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17806602#comment-17806602
]
Yu-Lin Chen commented on YUNIKORN-2325:
---------------------------------------
Attach the discussion history for how to calculate the top-N resources.
[https://yunikornworkspace.slack.com/archives/CL9CRJ1KM/p1701713212026479?thread_ts=1701680922.070739&cid=CL9CRJ1KM]
Craig Condit [1
个月前|https://yunikornworkspace.slack.com/archives/CL9CRJ1KM/p1701713212026479?thread_ts=1701680922.070739&cid=CL9CRJ1KM]
I think you can just look at the results. Take the count of each bucket for
each resource multiplied by the mean of that bucket's percentage (i.e. 0-10 =>
5, 10-20 -> 15, etc.) Whatever the top n resources are using that metric are
the ones we display.
Craig Condit [1
个月前|https://yunikornworkspace.slack.com/archives/CL9CRJ1KM/p1701713331408529?thread_ts=1701680922.070739&cid=CL9CRJ1KM]
So if you have 10 hosts at 90-100% (95%), and 5 hosts at 80-90% (85%), you
would use (10*0.95) + (5*0.85) = 13.75 for that resource. Each resource can be
calculated this way, and use the top n values (where n is number of bars you
want).
Craig Condit [1
个月前|https://yunikornworkspace.slack.com/archives/CL9CRJ1KM/p1701713341022139?thread_ts=1701680922.070739&cid=CL9CRJ1KM]
Just ignore the rest, as they aren't significant enough to matter.
Craig Condit [1
个月前|https://yunikornworkspace.slack.com/archives/CL9CRJ1KM/p1701713393468209?thread_ts=1701680922.070739&cid=CL9CRJ1KM]
The "dominant" resource could vary dramatically from one host to another. This
way, you show all the significant resources without making the graph difficult
to read.
Craig Condit [1
个月前|https://yunikornworkspace.slack.com/archives/CL9CRJ1KM/p1701713426559489?thread_ts=1701680922.070739&cid=CL9CRJ1KM]
And 10 resources might even be too much - maybe 5 is better... probably depends
on how the rendering looks. I'd experiment a bit. （已编辑）
Yu-Lin Chen [1
个月前|https://yunikornworkspace.slack.com/archives/CL9CRJ1KM/p1701715200174089?thread_ts=1701680922.070739&cid=CL9CRJ1KM]
Understood, my only concern is the domaint resource calculation differs from
the main node-utilization chart. (Which is calculated on core.) However,
since we're going to display top n resources, the result should have no
different. Will adopt the calculation, thanks [@Craig
Condit|https://yunikornworkspace.slack.com/team/U02CLPV3B5L]. （已编辑）
Wilfred Spiegelenburg [1
个月前|https://yunikornworkspace.slack.com/archives/CL9CRJ1KM/p1701742319324709?thread_ts=1701680922.070739&cid=CL9CRJ1KM]
This looks good,
For the number of bars, i.e. different resources, to show: I think 5 is good
number to settle on. When you look at the resources that would be dominant in a
cluster it will group around the same ones always: the most commonly requested.
In most clusters that would be memory, cpu, pods and GPU… As a fifth one
storage but that would probably be it. The fifth one is already not really
clear, showing 10 will be just cluttering.
> Add a chart to display multi-type resource utilisation (Web)
> ------------------------------------------------------------
>
> Key: YUNIKORN-2325
> URL: https://issues.apache.org/jira/browse/YUNIKORN-2325
> Project: Apache YuniKorn
> Issue Type: Sub-task
> Components: webapp
> Reporter: Yu-Lin Chen
> Assignee: Yu-Lin Chen
> Priority: Major
> Attachments: Multi-type resource utilisation chart (Draft).png
>
>
> Create a bar chart to display multi-type resource utilisation in YuniKorn UI
> Nodes page.
> The chart should:
> # Disaplay multi-type resources utilisation in YuniKorn Web UI Nodes page.
> # Calculate overall resource usage for each resource type and show top N
> resource (N=5)
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
