aurora-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Reza Motamedi <reza.motam...@gmail.com>
Subject Re: Review Request 66103: Introduce mesos disk collector
Date Thu, 22 Mar 2018 21:52:20 GMT

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/66103/
-----------------------------------------------------------

(Updated March 22, 2018, 9:52 p.m.)


Review request for Aurora, David McLaughlin, Daniel Knightly, Franck Cuny, Jordan Ly, Santhosh
Kumar Shanmugham, and Stephan Erb.


Changes
-------

- address feedback by adding timeout to `reuqests.get`.
- I also removed the `default_settings` class method in `DiskCollectorSettings` class since
it was essentially calling the constructor. Now instead of `DiskCollectorSettings.default_settings()`
I just call `DiskCollectorSettings()`. I don't think `default_settings` was very pythonic.


Repository: aurora


Description
-------

When disk isolation is enabled in a Mesos agent it calculates the disk usage for each container.

Thermos Observer also monitors disk usage using `twitter.common.dirutil`, essentially repeating
the work already done by the agent. In practice, we see that disk monitoring is one of the
most expensive resource monitoring tasks. For instance, when there are deeply nested directories,
the CPU utilization of the observer process can easily reach 1.5 CPUs. It would be ideal if
we delegate the disk monitoring task to the agent and do it only once. With this approach,
when disk collection has improved in the agent (for instance by implementing XFS isolation),
we can simply benefit from it without any code change. Some more information about the problem
is provided in AURORA-1918.

This patch that introduces `MesosDiskCollector` which queries the agent's API endpoint to
lookup disk_used_bytes. Note that there is also resource monitoring in thermos executor. Currently,
I left the disk collector there to use the `du` implementation. That can be changed in a later
patch.

I modified some vagrant config files including `aurora-executor.service` and `etc_mesos-slave/isolation`
for testing. They can be left as is. I included them in this patch to show how this would
work e2e.


Diffs (updated)
-----

  3rdparty/python/requirements.txt 4ac242cfa2c1c19cb7447816ab86e748839d3d11 
  RELEASE-NOTES.md 51ab6c724694244bf616b29e9beace4a4a3f5252 
  docs/reference/observer-configuration.md 8a443c94f7f37f9454989781f722101a97c99f15 
  examples/jobs/hello_world.aurora 5401bfebe753b5e53abd08baeac501144ced9b5a 
  examples/vagrant/mesos_config/etc_mesos-slave/isolation 1a7028ffc70116b104ef3ad22b7388f637707a0f

  examples/vagrant/systemd/thermos.service 01925bcd2ae44f100df511f3c3951c3f5a1a72aa 
  src/main/python/apache/aurora/tools/thermos_observer.py dd9f0c46ceac9e939b1b763073314161de0ea614

  src/main/python/apache/thermos/monitoring/BUILD 65ba7088f65e7baa5d30744736ba456b46a55e86

  src/main/python/apache/thermos/monitoring/disk.py 986d33a5000f8d5db15cb639c81f8b1d756ffa05

  src/main/python/apache/thermos/monitoring/resource.py adcdc751c03460dc801a18278faa96d6bd64722b

  src/main/python/apache/thermos/observer/task_observer.py a6870d48bddf2a2ccede7bb68195f2baae1d0e47

  src/test/python/apache/aurora/executor/common/test_resource_manager_integration.py fe74bd1d36666ecd89fca1b5b2251202cbbc0f24

  src/test/python/apache/thermos/monitoring/BUILD 8f2b39336dce6c7b580e6ba0009f60afdcb89179

  src/test/python/apache/thermos/monitoring/test_disk.py 362393bfd1facf3198e2d438d0596b16700b72b8

  src/test/python/apache/thermos/monitoring/test_resource.py e577e552d4ee1807096a15401851bb9fd95fa426



Diff: https://reviews.apache.org/r/66103/diff/7/

Changes: https://reviews.apache.org/r/66103/diff/6-7/


Testing
-------

- I added unit tests.
- Tested in vagrant and it works as intenced.
- I also built and deployed in our test enviroment. In order to measure imporoved performance
I created jobs with nested folders and noticed reduction in CPU utilization of the Observer
process, by at least 60%. (1.5 CPU cores to 0.4 CPU cores)

Here is one specific test setup: On two hosts I created a two tasks. Each task creates identical
nested directory structures and files in them. The overall size is 30GB. test_host_1 runs
the current version of observer and test_host_2 runs Observer with this patch and also has
mesos_disk_collection enabled. The results are as follows:

```
rezam[7]TEST_HOST_1 ~ $ while true; do echo `date`; curl localhost:1338/vars -s | grep cpu;
sleep 10; done
Thu Mar 22 04:36:17 UTC 2018
observer.observer_cpu 108.9
Thu Mar 22 04:36:27 UTC 2018
observer.observer_cpu 123.2
Thu Mar 22 04:36:38 UTC 2018
observer.observer_cpu 123.2
Thu Mar 22 04:36:48 UTC 2018
observer.observer_cpu 123.2
Thu Mar 22 04:36:58 UTC 2018
observer.observer_cpu 111.0
Thu Mar 22 04:37:08 UTC 2018
observer.observer_cpu 111.0
Thu Mar 22 04:37:18 UTC 2018
observer.observer_cpu 111.0


rezam[7]TEST_HOST_2 ~ $ while true; do echo `date`; curl localhost:1338/vars -s | grep cpu;
sleep 10; done
Thu Mar 22 04:36:20 UTC 2018
observer.observer_cpu 1.3
Thu Mar 22 04:36:30 UTC 2018
observer.observer_cpu 1.3
Thu Mar 22 04:36:40 UTC 2018
observer.observer_cpu 1.3
Thu Mar 22 04:36:50 UTC 2018
observer.observer_cpu 1.2
Thu Mar 22 04:37:00 UTC 2018
observer.observer_cpu 1.2
Thu Mar 22 04:37:10 UTC 2018
observer.observer_cpu 1.2
Thu Mar 22 04:37:20 UTC 2018
observer.observer_cpu 1.8
```


Thanks,

Reza Motamedi


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