> On Jan. 21, 2016, 1:20 a.m., Klaus Ma wrote:
> > src/master/allocator/mesos/hierarchical.cpp, line 1313
> > <https://reviews.apache.org/r/42559/diff/1/?file=1203488#file1203488line1313>
> >
> >     Who will kill this `revocable` resources? So we'll have a patch for 
> > evicting?
> 
> Alexander Rukletsov wrote:
>     Someday, someway... : )
>     
>     Actually, I thought you guys are working on this (oversubscription for 
> reservations, incl. quota)!

In "oversubscription for reservations", only `reserved & !persistentVolume` 
resources are considered to be revocable to other framework; the un-reserved 
resources are not handled for now.

To me, the key of eviction is which resources shoulbe be evicted for its owner: 
1.) in "oversubscription for reservations", it's `reserved & 
!persistentVolume`, 2.) in "oversubscription" (Estimator & QoSController), it 
dependent on modules;

When I polishing the design doc of Phase 2 (Simultaneous offer), I'm thinking 
to do fair-share un-reserved resources by Quota/Role's Weight; for example, 100 
CPUS in cluster and two Quota: 30 CPUS (role1) & 20 CPUS (role2); when only 
framework1 in role1 registered, it will get 30 CPUS (reserved) + 30 CPUS 
(unreserved) + 20 CPUS (revocable), this offer is Quota + (total - Quota1 - 
Quota2) * (Quota1/(Quota1 + Quota2)) + (total - Quota1 - Quota2) * 
(Quota2/(Quota1 + Quota2)); when framework2 in role2 registered, the revocable 
20CPUS of framework1 will be evicted by master/allocator.

I logged MESOS-4303 to trace this idea; but did not get enough time on the 
detail for now :). Maybe we can draft a doc on this, any comments?


- Klaus


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/42559/#review115443
-----------------------------------------------------------


On Jan. 21, 2016, 1:01 a.m., Alexander Rukletsov wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/42559/
> -----------------------------------------------------------
> 
> (Updated Jan. 21, 2016, 1:01 a.m.)
> 
> 
> Review request for mesos, Bernd Mathiske, Joris Van Remoortere, Joseph Wu, 
> and Qian Zhang.
> 
> 
> Bugs: MESOS-4441
>     https://issues.apache.org/jira/browse/MESOS-4441
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> See summary.
> 
> 
> Diffs
> -----
> 
>   docs/quota.md 1a6d2f07fb74d168a7eb30764ab9ff80cea5e3b6 
>   src/master/allocator/mesos/hierarchical.cpp 
> e32ee4aa3ed9793bb5a99233e699e5cc2bdd796b 
>   src/tests/hierarchical_allocator_tests.cpp 
> 953712149bd951789beb29c72779c4ac65aa48dc 
> 
> Diff: https://reviews.apache.org/r/42559/diff/
> 
> 
> Testing
> -------
> 
> On Mac OS 10.10.4:
> 
> `GTEST_FILTER="HierarchicalAllocatorTest.*" ./bin/mesos-tests.sh 
> --gtest_repeat=100 --gtest_break_on_failure --gtest_shuffle`
> `make check`
> 
> Updated `HierarchicalAllocatorTest.DRFWithQuota` succeeds with the change and 
> fails without.
> 
> 
> Thanks,
> 
> Alexander Rukletsov
> 
>

Reply via email to