[
https://issues.apache.org/jira/browse/YARN-4080?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14711622#comment-14711622
]
MENG DING commented on YARN-4080:
---------------------------------
Not sure if the title accurately reflects the problem. If you think there is a
better way to describe the problem, please suggest.
For the use case presented in the description, one possible direction to
consider is something like a dynamic host-based reservation (note, this is not
the same as the current container reservation in YARN), for example:
* when asking for resource requirement, one can specify the initial resource
capability, and a reserved resource capability on whatever host that the
container is launched on. For example, I can say I want 2GB of initial resource
for a container, and once that container is launched, reserve up to 16GB of
resource for the container on that host, as I expect the resource usage of the
container will fluctuate over time, and will sometime peak at 16GB.
* if this reserved resource is not fully utilized, it can still be allocated to
other applications, but the scheduler will indicate that the allocated resource
is revocable, such that no critical service should use this chunk of resource
* when scheduler is allocating new resource, it should first consider resource
that has not been reserved
* preemption logic should also preempt these kind of revocable resource if
needed
The above is similar to the dynamic reservation feature being implemented in
Mesos: https://issues.apache.org/jira/browse/MESOS-2018
I also took a look at YARN-1051 to see if the current reservation system in
YARN could help with this situation, but to the best of my knowledge, it seems
to mainly address applications with a future start time and a predictable
deadline. Please correct me if I am wrong.
Let me know if you have any thoughts, comments or ideas.
> Capacity planning for long running services on YARN
> ---------------------------------------------------
>
> Key: YARN-4080
> URL: https://issues.apache.org/jira/browse/YARN-4080
> Project: Hadoop YARN
> Issue Type: Improvement
> Components: api, resourcemanager
> Reporter: MENG DING
>
> YARN-1197 addresses the functionality of container resource resize. One major
> use case of this feature is for long running services managed by Slider to
> dynamically flex up and down resource allocation of individual components
> (e.g., HBase region server), based on application metrics/alerts obtained
> through third-party monitoring and policy engine.
> One key issue with increasing container resource at any point of time is that
> the additional resource needed by the application component may not be
> available *on the specific node*. In this case, we need to rely on preemption
> logic to reclaim the required resource back from other (preemptable)
> applications running on the same node. But this may not be possible today
> because:
> * preemption doesn't consider constraints of pending resource requests, such
> as hard locality requirements, user limits, etc (being addressed in YARN-2154
> and possibly in YARN-3769?)
> * there may not be any preemptable container available due to the fact that
> no application is over its guaranteed capacity.
> What we need, ideally, is a way for YARN to support future capacity planning
> of long running services. At the minimum, we need to provide a way to let
> YARN know about the resource usage prediction/pattern of a long running
> service. And given this knowledge, YARN should be able to preempt resources
> from other applications to accommodate the resource needs of the long running
> service.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
