You are viewing a plain text version of this content. The canonical link for it is here.

Posted to yarn-dev@hadoop.apache.org by Arun Suresh <as...@apache.org> on 2018/01/25 23:46:39 UTC

[DISCUSS] Merge YARN-6592 to trunk

Hello yarn-dev@

We feel that the YARN-6592 dev branch mostly in shape to be merged into
trunk. This branch adds support for placing containers in YARN using rich
placement constraints. For example, this can be used by applications to
co-locate containers on a node or rack (*affinity *constraint), spread
containers across nodes or racks (*anti-affinity* constraint), or even
specify the maximum number of containers on a node/rack (*cardinality *
constraint).

We have integrated this feature into the Distributed-Shell application for
feature testing. We have performed end-to-end testing on moderately-sized
clusters to verify that constraints work fine. Performance tests have been
done via both SLS tests and Capacity Scheduler performance unit tests, and
no regression was found. We have opened a JIRA to track Jenkins acceptance
of the aggregated patch [2]. Documentation is in the process of being
completed [3]. You can also check our design document for more details [4].

Config flags are needed to enable this feature and it should not have any
effect on YARN when turned off. Once merged, we plan to work on further
improvements, which can be tracked in the umbrella YARN-7812 [5].

Kindly do take a look at the branch and raise issues/concerns that need to
be addressed before the merge.

Many thanks to Konstantinos, Wangda, Panagiotis, Weiwei, and Sunil for
their contributions to this effort, as well as Subru, Chris, Carlo, and
Vinod for their inputs and discussions.

Cheers
-Arun


[1] https://issues.apache.org/jira/browse/YARN-6592
[2] https://issues.apache.org/jira/browse/YARN-7792
[3] https://issues.apache.org/jira/browse/YARN-7780
[4]
https://issues.apache.org/jira/secure/attachment/12867869/YARN-6592-Rich-Placement-Constraints-Design-V1.pdf
[5] https://issues.apache.org/jira/browse/YARN-7812

UNCHECKED Re: [DISCUSS] Merge YARN-6592 to trunk

Posted by Arun Suresh <ar...@gmail.com>.

Thanks everyone. I've opened a VOTE thread..

@Carlo
Interesting idea to support Kubernates API. Definitely makes sense. My
understanding is that we can emulate specification for a k8 POD within our
scheme of things by assigning the same source tag to a set of scheduling
requests.
Ill dig in to this.

Cheers
-Arun

On Fri, Jan 26, 2018 at 1:34 AM, Carlo Aldo Curino <ca...@gmail.com>
wrote:

> Also I think this can help us close the gap (and surpass) Kubernetes for
> complex services (at least for resource management)... It would be awesome
> to have a compatibility layer so folks can run Kubernetes natives apps on a
> yarn cluster.
>
>
>
>
> On Jan 26, 2018 1:32 AM, "Carlo Aldo Curino" <ca...@gmail.com>
> wrote:
>
> +1. I didn't runs tests, but I like the design, and speaking with ops
> teams that operate large clusters I hear this is a feature they think is
> going to help a lot, so I am very supportive of this effort.
>
> On Jan 25, 2018 7:08 PM, "Konstantinos Karanasos" <kk...@gmail.com>
> wrote:
>
>> Thanks for starting the thread Arun, +1 from me too.
>>
>> Konstantinos
>>
>> On Thu, Jan 25, 2018 at 18:54 Weiwei Yang <ch...@hotmail.com> wrote:
>>
>>> +1, thanks for getting to this milestone Arun.
>>> I’ve done some basic validations on a 4 nodes cluster, with some general
>>> affinity/anti-affinty/cardinality constraints, it worked. I’ve also
>>> reviewed the doc, it’s in good shape and very illustrative.
>>>
>>> Thanks.
>>>
>>> --
>>> Weiwei
>>>
>>> On 26 Jan 2018, 10:44 AM +0800, Sunil G <su...@apache.org>, wrote:
>>> +1.
>>>
>>> Thanks Arun.
>>>
>>> I did manual testing for check affinity and anti-affinity features with
>>> placement allocator. Also checked SLS to see any performance regression,
>>> and there are not much difference as Arun mentioned.
>>>
>>> Thanks all the folks for working on this. Kudos!
>>>
>>> - Sunil
>>>
>>>
>>> On Fri, Jan 26, 2018 at 5:16 AM Arun Suresh <asuresh@apache.org<mailto:
>>> asuresh@apache.org>> wrote:
>>> Hello yarn-dev@
>>>
>>> We feel that the YARN-6592 dev branch mostly in shape to be merged into
>>> trunk. This branch adds support for placing containers in YARN using rich
>>> placement constraints. For example, this can be used by applications to
>>> co-locate containers on a node or rack (affinity constraint), spread
>>> containers across nodes or racks (anti-affinity constraint), or even
>>> specify the maximum number of containers on a node/rack (cardinality
>>> constraint).
>>>
>>> We have integrated this feature into the Distributed-Shell application
>>> for feature testing. We have performed end-to-end testing on
>>> moderately-sized clusters to verify that constraints work fine. Performance
>>> tests have been done via both SLS tests and Capacity Scheduler performance
>>> unit tests, and no regression was found. We have opened a JIRA to track
>>> Jenkins acceptance of the aggregated patch [2]. Documentation is in the
>>> process of being completed [3]. You can also check our design document for
>>> more details [4].
>>>
>>> Config flags are needed to enable this feature and it should not have
>>> any effect on YARN when turned off. Once merged, we plan to work on further
>>> improvements, which can be tracked in the umbrella YARN-7812 [5].
>>>
>>> Kindly do take a look at the branch and raise issues/concerns that need
>>> to be addressed before the merge.
>>>
>>> Many thanks to Konstantinos, Wangda, Panagiotis, Weiwei, and Sunil for
>>> their contributions to this effort, as well as Subru, Chris, Carlo, and
>>> Vinod for their inputs and discussions.
>>>
>>> Cheers
>>> -Arun
>>>
>>>
>>> [1] https://issues.apache.org/jira/browse/YARN-6592
>>> [2] https://issues.apache.org/jira/browse/YARN-7792
>>> [3] https://issues.apache.org/jira/browse/YARN-7780
>>> [4] https://issues.apache.org/jira/secure/attachment/12867869/YA
>>> RN-6592-Rich-Placement-Constraints-Design-V1.pdf
>>> [5] https://issues.apache.org/jira/browse/YARN-7812
>>>
>>> --
>> Konstantinos
>>
>
>

UNCHECKED Re: [DISCUSS] Merge YARN-6592 to trunk

Posted by Arun Suresh <ar...@gmail.com>.

Thanks everyone. I've opened a VOTE thread..

@Carlo
Interesting idea to support Kubernates API. Definitely makes sense. My
understanding is that we can emulate specification for a k8 POD within our
scheme of things by assigning the same source tag to a set of scheduling
requests.
Ill dig in to this.

Cheers
-Arun

On Fri, Jan 26, 2018 at 1:34 AM, Carlo Aldo Curino <ca...@gmail.com>
wrote:

> Also I think this can help us close the gap (and surpass) Kubernetes for
> complex services (at least for resource management)... It would be awesome
> to have a compatibility layer so folks can run Kubernetes natives apps on a
> yarn cluster.
>
>
>
>
> On Jan 26, 2018 1:32 AM, "Carlo Aldo Curino" <ca...@gmail.com>
> wrote:
>
> +1. I didn't runs tests, but I like the design, and speaking with ops
> teams that operate large clusters I hear this is a feature they think is
> going to help a lot, so I am very supportive of this effort.
>
> On Jan 25, 2018 7:08 PM, "Konstantinos Karanasos" <kk...@gmail.com>
> wrote:
>
>> Thanks for starting the thread Arun, +1 from me too.
>>
>> Konstantinos
>>
>> On Thu, Jan 25, 2018 at 18:54 Weiwei Yang <ch...@hotmail.com> wrote:
>>
>>> +1, thanks for getting to this milestone Arun.
>>> I’ve done some basic validations on a 4 nodes cluster, with some general
>>> affinity/anti-affinty/cardinality constraints, it worked. I’ve also
>>> reviewed the doc, it’s in good shape and very illustrative.
>>>
>>> Thanks.
>>>
>>> --
>>> Weiwei
>>>
>>> On 26 Jan 2018, 10:44 AM +0800, Sunil G <su...@apache.org>, wrote:
>>> +1.
>>>
>>> Thanks Arun.
>>>
>>> I did manual testing for check affinity and anti-affinity features with
>>> placement allocator. Also checked SLS to see any performance regression,
>>> and there are not much difference as Arun mentioned.
>>>
>>> Thanks all the folks for working on this. Kudos!
>>>
>>> - Sunil
>>>
>>>
>>> On Fri, Jan 26, 2018 at 5:16 AM Arun Suresh <asuresh@apache.org<mailto:
>>> asuresh@apache.org>> wrote:
>>> Hello yarn-dev@
>>>
>>> We feel that the YARN-6592 dev branch mostly in shape to be merged into
>>> trunk. This branch adds support for placing containers in YARN using rich
>>> placement constraints. For example, this can be used by applications to
>>> co-locate containers on a node or rack (affinity constraint), spread
>>> containers across nodes or racks (anti-affinity constraint), or even
>>> specify the maximum number of containers on a node/rack (cardinality
>>> constraint).
>>>
>>> We have integrated this feature into the Distributed-Shell application
>>> for feature testing. We have performed end-to-end testing on
>>> moderately-sized clusters to verify that constraints work fine. Performance
>>> tests have been done via both SLS tests and Capacity Scheduler performance
>>> unit tests, and no regression was found. We have opened a JIRA to track
>>> Jenkins acceptance of the aggregated patch [2]. Documentation is in the
>>> process of being completed [3]. You can also check our design document for
>>> more details [4].
>>>
>>> Config flags are needed to enable this feature and it should not have
>>> any effect on YARN when turned off. Once merged, we plan to work on further
>>> improvements, which can be tracked in the umbrella YARN-7812 [5].
>>>
>>> Kindly do take a look at the branch and raise issues/concerns that need
>>> to be addressed before the merge.
>>>
>>> Many thanks to Konstantinos, Wangda, Panagiotis, Weiwei, and Sunil for
>>> their contributions to this effort, as well as Subru, Chris, Carlo, and
>>> Vinod for their inputs and discussions.
>>>
>>> Cheers
>>> -Arun
>>>
>>>
>>> [1] https://issues.apache.org/jira/browse/YARN-6592
>>> [2] https://issues.apache.org/jira/browse/YARN-7792
>>> [3] https://issues.apache.org/jira/browse/YARN-7780
>>> [4] https://issues.apache.org/jira/secure/attachment/12867869/YA
>>> RN-6592-Rich-Placement-Constraints-Design-V1.pdf
>>> [5] https://issues.apache.org/jira/browse/YARN-7812
>>>
>>> --
>> Konstantinos
>>
>
>