mesos-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "James Peach" <jpe...@apache.org>
Subject Re: Review Request 37873: Add quiesce logic in allocator
Date Mon, 31 Aug 2015 17:46:26 GMT


> On Aug. 30, 2015, 5:10 p.m., Robert Lacroix wrote:
> > src/master/allocator/mesos/hierarchical.hpp, line 909
> > <https://reviews.apache.org/r/37873/diff/3/?file=1059860#file1059860line909>
> >
> >     We need to store the information that a framework is quiescing offers in a more
memory efficient way. This doesn't work for clusters with a lot of frameworks and a lot of
slaves.
> >     
> >     Can't we just use a boolean to indicate the framework has quiesced?
> 
> Guangya Liu wrote:
>     The reason that I did not use a boolean value is because here quiesceOffers is a
complement to the current reviveOffers, and reviveOffers can just clear the filters to enable
the framework can get resource again, so here I was setting filters to this framework to disable
it getting offers. For performance issue, the current settting are all memory operations and
only for framework, I think that the Mesos cluster may not have thousands of frameworks. Comments?
Thanks.
> 
> Robert Lacroix wrote:
>     reviveOffers could just flip the boolean in addition to removing the filters, so
that shouldn't make the logic much more complicated.
>     
>     The number of filters is indeed a problem because they (and all of their resources
- which in case the ports are fragmented can be a lot) need to be traversed on every allocation.
> 
> Guangya Liu wrote:
>     Thanks Robert for the update. The design for quiesce offer also want a timeout for
quiesce, if the timeout reached, then the framework can get resource offers automatically
and no need for cluster admin to get involved to revive offer. Do you have any comments for
adding a filter to quiesce offer? Is there are any other methods to add a filter which is
more efficient? Hope can get some comments from you.

You can add the quiesce timeout using the mechanism that refuse timeouts use. It might get
a little more involved since if you receive a second quiesce, you would need to update the
timeout. You can't implement quiesce using a filter because filters are not ordered. You have
to be able to check whether a framework is quiesced without traversing the whole filter set.


- James


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/37873/#review97031
-----------------------------------------------------------


On Aug. 31, 2015, 5:49 a.m., Guangya Liu wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/37873/
> -----------------------------------------------------------
> 
> (Updated Aug. 31, 2015, 5:49 a.m.)
> 
> 
> Review request for mesos, Ben Mahler, Jie Yu, and Vinod Kone.
> 
> 
> Bugs: MESOS-3037
>     https://issues.apache.org/jira/browse/MESOS-3037
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> Add quiesce logic in allocator
> 
> 
> Diffs
> -----
> 
>   include/mesos/master/allocator.hpp 659f37b3f9d9fa02da9bdb6c85cd3c180a24b73a 
>   src/master/allocator/mesos/allocator.hpp aa55755a9c3250579e9366bdbc17a2449e95d659 
>   src/master/allocator/mesos/hierarchical.hpp 38f8fd2c84314bb3731684d0e9795cb4f50a227e

>   src/master/master.cpp 95207d24db0aa052eb70c4cc7eb75d0611c365cf 
>   src/tests/mesos.hpp b2160f50db3e0b0d04b0625e7589004016c0f746 
> 
> Diff: https://reviews.apache.org/r/37873/diff/
> 
> 
> Testing
> -------
> 
> 
> Thanks,
> 
> Guangya Liu
> 
>


Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message