mesos-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Guangya Liu <gyliu...@gmail.com>
Subject Re: Review Request 49784: Increase framework numbers to allocator benchmarks.
Date Fri, 15 Jul 2016 22:52:58 GMT


> On 七月 12, 2016, 4:32 a.m., Guangya Liu wrote:
> > src/tests/hierarchical_allocator_tests.cpp, line 3196
> > <https://reviews.apache.org/r/49784/diff/1/?file=1439341#file1439341line3196>
> >
> >     Two issues for this patch:
> >     
> >     1) The new added framework number will introduce some cases that agent number
is less than framework number and some frameworks may not able to get resources.
> >     2) This will slow down the unit test, it is better introduce this and fix https://issues.apache.org/jira/browse/MESOS-4558
together by introducing `batchsize` for benchmark test.
> 
> Alexander Rukletsov wrote:
>     1) I think this a good thing to test / benchmark.
>     2) Benchmark tests are not part of the default test suite, hence it seems fine if
this test takes more time to run.
> 
> Guangya Liu wrote:
>     For 2), one minor comment is MESOS-4558 is tracking the time of benchmark tests because
want to enable them on ASF CI, so it is better to reduce the time of the benchmark test.
> 
> Jiang Yan Xu wrote:
>     I commented on MESOS-4558. For this patch I think this is fine because speeding up
benchmark tests (or just CI) isn't necessarily at odds with increasing the cluster size. The
motivation for the latter is that we do have this many frameworks in our cluster and we'd
like to test it.
> 
> Jiang Yan Xu wrote:
>     About 1), I don't think in the tests we have the expectation that all frameworks
need to be allocated resources in each allocation round? Also this scenario is very much real
for us. :)

Thanks Jiang Yan, before the fix, I saw the max framework count is same as the minimum agent
count, so my thinking is that we always want all of the frameworks get resources in the benchmark
test. But after some performance test recently, I found that this is useful for the evaluating
the case `how much time does the fully allocated agent contribute to the allocation cycle`,
i.e. 3000 frameworks + 1000 agents, 6000 frameworks + 1000 agents etc.

```
// The Hierarchical Allocator benchmark tests are parameterized
// by the number of slaves.
INSTANTIATE_TEST_CASE_P(
  SlaveAndFrameworkCount,
  HierarchicalAllocator_BENCHMARK_Test,
  ::testing::Combine(
    ::testing::Values(1000U, 5000U, 10000U, 20000U, 30000U, 50000U),
    ::testing::Values(1U, 50U, 100U, 200U, 500U, 1000U, 3000U, 6000U))
  );
```


- Guangya


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/49784/#review141843
-----------------------------------------------------------


On 七月 8, 2016, 4:02 a.m., Jacob Janco wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/49784/
> -----------------------------------------------------------
> 
> (Updated 七月 8, 2016, 4:02 a.m.)
> 
> 
> Review request for mesos and Jiang Yan Xu.
> 
> 
> Bugs: MESOS-5780 and MESOS-5781
>     https://issues.apache.org/jira/browse/MESOS-5780
>     https://issues.apache.org/jira/browse/MESOS-5781
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> - Our clusters have very high numbers of frameworks,
>   and we would like to increase the allocator benchmark
>   parameters to reflect this.
> 
> 
> Diffs
> -----
> 
>   src/tests/hierarchical_allocator_tests.cpp 0498cd5e54b0e4b87a767585a77699653aa52179

> 
> Diff: https://reviews.apache.org/r/49784/diff/
> 
> 
> Testing
> -------
> 
> make check
> 
> 
> Thanks,
> 
> Jacob Janco
> 
>


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