[
https://issues.apache.org/jira/browse/YARN-462?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13601245#comment-13601245
]
Kendall Thrapp commented on YARN-462:
-------------------------------------
Thanks for the questions and feedback. Yes, first I should clarify what I
intended by chargeback. I'm looking to be able quantify cluster resource usage
(memory, CPU, HDFS, etc.) for every application, and then roll that up to the
project level. This would allow us to accurately charge the customer (i.e.
team/project) for their grid usage (either literally or just informatively). I
want to provide incentive for more efficient coding, as well as make it easier
for teams to compare their resource usage across different software versions of
their Hadoop applications, config parameter changes, etc.
I had originally hoped that hierarchical queues could serve this purpose as
well, but have since run into several issues with this approach. The first is
that it doesn't scale for clusters with large numbers of projects. I've seen
large clusters shared between over a hundred different projects, each with
their own teams of users. If I recall correctly, queues can't be assigned less
than 1% of the total capacity, so it wouldn't be possible to give each of these
project their own queue. Even if we could, I suspect this could result in too
much overhead for the scheduler and too much fragmentation of the cluster
resources, which could result in poorer overall utilization.
The second issue is that the project-per-queue approach conflicts with how I
see users wanting to use our queues. In many cases I see queues being used to
distinguish application priorities, ensuring that high priority time-sensitive
jobs get the resources they need to finish on time, while big but lower
priority and less time-sensitive jobs are constrained by being in a smaller
queue. I'd expect a lot of pushback from our users for any chargeback-focused
queue configuration that had a negative impact on job run times and meeting
SLAs. The idea of the project/chargeback parameter decouples the two.
> Project Parameter for Chargeback
> --------------------------------
>
> Key: YARN-462
> URL: https://issues.apache.org/jira/browse/YARN-462
> Project: Hadoop YARN
> Issue Type: New Feature
> Components: resourcemanager
> Affects Versions: 0.23.6
> Reporter: Kendall Thrapp
>
> Problem Summary
> For the purpose of chargeback and better understanding of grid usage, we need
> to be able to associate applications with "projects", e.g. "pipeline X",
> "property Y". This would allow us to aggregate on this property, thereby
> helping us compute grid resource usage for the entire "project". Currently,
> for a given application, two things we know about it are the user that
> submitted it and the queue it was submitted to. Below, I'll explain why
> neither of these is adequate for enterprise-level chargeback and
> understanding resource allocation needs.
> Why Not Users?
> Its not individual users that are paying the bill -- its projects. When one
> of our real users submits an application on a Hadoop grid, they're presumably
> not usually doing it for themselves. They're doing work for some project or
> team effort, so its that team or project that should be "charged" for all its
> users applications. Maintaining outside lists of associations between users
> and projects is error-prone because it is time-sensitive and requires
> continued ongoing maintenance. New users join organizations, users leave and
> users even change projects. Furthermore, users may split their time between
> multiple projects, making it ambiguous as to which of a user's projects a
> given application should be charged. Also, there can be headless users,
> which can be even more difficult to link to a project and can be shared
> between teams or projects.
> Why Not Queues?
> The purpose of queues is for scheduling. Overloading the queues concept to
> also mean who should be "charged" for an application can have a detrimental
> effect on the primary purpose of queues. It could be manageable in the case
> of a very small number of projects sharing a cluster, but doesn't scale to
> tens or hundreds of projects sharing a cluster. If a given cluster is shared
> between 50 projects, creating 50 separate queues will result in inefficient
> use of the cluster resources. Furthermore, a given project may desire more
> than one queue for different types or priorities of applications.
> Proposed Solution
> Rather than relying on external tools to infer through the user and/or queue
> who to "charge" for a given application, I propose a straightforward approach
> where that information be explicitly supplied when the application is
> submitted, just like we do with queues. Let's use a charge card analogy:
> when you buy something online, you don't just say who you are and how to ship
> it, you also specify how you're paying for it. Similarly, when submitting an
> application in YARN, you could explicitly specify to whom it's resource usage
> should be associated (a project, team, cost center, etc).
> This new configuration parameter should default to being optional, so that
> organizations not interested in chargeback or project-level resource tracking
> can happily continue on as if it wasn't there. However, it should be
> configurable at the cluster-level such that, a given cluster to could elect
> to make it required, so that all applications would have an associated
> project. The value of this new parameter should be exposed via the Resource
> Manager UI and Resource Manager REST API, so that users and tools can make
> use of it for chargeback, utilization metrics, etc.
> I'm undecided on what to name the new parameter, as I like the flexibility in
> the ways it could be used. It is essentially just an additional party other
> than user or queue that an application can be associated with, so its use is
> not just limited to a chargeback scenario. For example, an organization not
> interested in chargeback could still use this parameter to communicate useful
> information about a application (e.g. pipelineX.stageN) and aggregate like
> applications.
> Enforcement
> Couldn't users just specify this information as a prefix for their job names?
> Yes, but the missing piece this could provides is enforcement. Ideally, I'd
> like this parameter to work very much like how the queues work. Like already
> exists with queues, it'd be ideal if a given user couldn't just specify any
> old value for this parameter. It could be configurable such that a given
> user only has permission to submit applications for specific "projects".
> Submitting an application with this parameter being anything other than what
> the given user is allowed, would cause the application to be rejected in the
> same manner as if the user has specified an invalid queue.
> Again, so as to have no effect on organizations not interested in this
> feature, this enforcement should be off by default, but configurable at the
> cluster level such that it could be turned on for clusters wanting to use it.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
