[jira] [Comment Edited] (MESOS-4738) Expose egress bandwidth as a resource

["Avinash Sridharan (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MESOS-4738?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15163362#comment-15163362 ] 

Avinash Sridharan edited comment on MESOS-4738 at 2/24/16 5:30 PM:
-------------------------------------------------------------------

I agree with [~w013ccw] on this, we don't have a clean way of putting ingress bandwidth limits and hence should focus on egress. As far as treating egress bandwidth as a resource, we can treat this resource as a minimal rate guarantee rather than a fixed rate allocation. Using a DRR scheduler (https://en.wikipedia.org/wiki/Deficit_round_robin) on the egress rates should guarantee a minimum rate to each container . I think qdisc does support DRR (http://manpages.ubuntu.com/manpages/raring/man8/tc-drr.8.html) .


was (Author: avinash@mesosphere.io):
I agree with [~w013ccw] on this, we don't have a clean way of putting ingress bandwidth limits and hence should focus on egress. As far as treating egress bandwidth as a resource, we can treat this resource as a minimal rate guarantee rather than a fixed rate allocation. Using a DRR scheduler (https://en.wikipedia.org/wiki/Deficit_round_robin) on the egress rates should guarantee a minimum rate to each container . I think qdisc does support DRR.

> Expose egress bandwidth as a resource
> -------------------------------------
>
>                 Key: MESOS-4738
>                 URL: https://issues.apache.org/jira/browse/MESOS-4738
>             Project: Mesos
>          Issue Type: Improvement
>            Reporter: Sargun Dhillon
>            Priority: Minor
>              Labels: mesosphere
>
> Some of our users care about variable network network isolation. Although we cannot fundamentally limit ingress network bandwidth, having it as a resource, so we can drop packets above a specific limit would be attractive. 
> It would be nice to expose egress and ingress bandwidth as an agent resource, perhaps with a default of 10,000 mbps, and we can allow people to adjust as needed. Alternatively, a more advanced design would involve generating heuristics based on an analysis of the network MII / PHY. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)