[
https://issues.apache.org/jira/browse/YARN-4039?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14694042#comment-14694042
]
Tsuyoshi Ozawa commented on YARN-4039:
--------------------------------------
[~frsyuki] thank you for taking this issue. The problem should be fixed since
it can cause unexpected starvation by resource allocations for Application
Masters. There are 2 solutions to address this problem:
1. Supporting gang scheduling at YARN-level natively.
As [~tucu00]
[mentioned|https://issues.apache.org/jira/browse/YARN-624?focusedCommentId=13658637&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13658637],
this scheduling method can solve the resource starvation because lots AMs
consume containers unexpectedly.
As a workaround implementation for supporting gang scheduling for small jobs,
one possible way is to fall back to uber mode when job is small enough. It
prevents deadlock because of AM allocations. However, some problems are under
discussion.
2. Changing the priority of AMs before launching containers for *tasks* as you
uploaded. This way can solve the problem, but there is a tradeoff as you
mentioned.
IMHO, the approach 2 is more easy to implement. Do you have any result of
benchmark for making the influences of the patch clear?
Additionally, minor nits at code level:
{code}
+ if (weight.getWeight(ResourceType.MEMORY) <
targetWeight.getWeight(ResourceType.MEMORY)) {
{code}
Could you update to use ResourceCalculator#compare via FSQueue#SchedulingPolicy
instead of getWeight() because it can break semantics of DRF? Also, could you
update the patch on trunk code?
> New AM instances waste resource by waiting only for resource availability
> when all available resources are already used
> -----------------------------------------------------------------------------------------------------------------------
>
> Key: YARN-4039
> URL: https://issues.apache.org/jira/browse/YARN-4039
> Project: Hadoop YARN
> Issue Type: Improvement
> Components: fairscheduler
> Affects Versions: 2.4.0, 2.5.0, 2.6.0, 2.7.0
> Reporter: Sadayuki Furuhashi
> Assignee: Sadayuki Furuhashi
> Attachments: YARN-4039.1.patch
>
>
> Problem:
> In FairScheduler, maxRunningApps doesn't work well if we can't predict size
> of an application in a queue because small maxRunningApps can't use all
> resources if many small applications are issued, where large maxRunningApps
> wastes resources if large applications run.
> Background:
> We're using FairScheduler. In following scenario, AM instances wastes
> resources significantly:
> * A queue has X MB of capacity.
> * An application requests 32 containers where a container requires (X / 32)
> MB of memory
> ** In this case, a single application occupies entire resource of the queue.
> * Many those applications are issued (10 applications)
> * Ideal behavior is that applications run one by one to maximize throughput.
> * However, all applications run simultaneously. As the result, AM instances
> occupy resources and prevent other tasks from starting. At worst case, most
> of resources are occupied by waiting AMs and applications progress very
> slowly.
> A solution is setting maxRunningApps to 1 or 2. However, it doesn't work well
> if following workload exists at the same queue:
> * An application requests 2 containers where a container requires (X / 32) MB
> of memory
> * Many those applications are issued (say, 10 applications)
> * Ideal behavior is that all applications run simultaneously to maximize
> concurrency and throughput.
> * However, number of applications are limited by maxRunningApps. At worst
> case, most of resources are idling.
> This problem happens especially with Hive because we can't estimate size of a
> MapReduce application.
> Solution:
> AM doesn't have to start if there are waiting resource requests because the
> AM can't grant resource requests even if it starts.
> Patch:
> I attached a patch that implements this behavior. But this implementation has
> this trade-off:
> * When AM is registered to FairScheduler, its demand is 0 because even AM
> attempt is not created. Starting this AM doesn't change resource demand of a
> queue. So, if many AMs are issued to a queue at the same time, all AMs will
> be RUNNING. But we want to prevent it.
> * When a AM starts, demand of the AM is only AM attempt. Then AM requires
> more resources. Until AM requires resources, demand of the queue is low. But
> starting AM during this time will start unnecessary AMs.
> * So, this patch doesn't start immediately when AM is registered. Instead, it
> starts AM only every continuous-scheduling-sleep-ms.
> * Setting large continuous-scheduling-sleep-ms will prevent wasting AMs. But
> increases latency.
> Therefore, this patch is enabled only if new option "demand-block-am-enabled"
> is true.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
