David Robinson created AURORA-1918:
--------------------------------------
Summary: allow resource monitoring to be disabled in the executor
Key: AURORA-1918
URL: https://issues.apache.org/jira/browse/AURORA-1918
Project: Aurora
Issue Type: Task
Components: Executor
Reporter: David Robinson
Assignee: David Robinson
The Aurora executor monitors a [task's resource
usage|https://github.com/apache/aurora/blob/cc2aa46f7ad8590e201621ffe2799299959ef7eb/src/main/python/apache/thermos/monitoring/resource.py#L15-L28]
(CPU, memory and disk) and kills it [if its disk usage exceeds its
reservation|https://github.com/apache/aurora/blob/cc2aa46f7ad8590e201621ffe2799299959ef7eb/src/main/python/apache/aurora/executor/common/resource_manager.py#L61-L67].
Monitoring disk usage is expensive, the executor does the equivalent of running
'du' inside a container sandbox; it recursively walks the sandbox to calculate
usage and in doing so effectively trashes the page cache. Within Twitter we've
seen the executor consume an entire core while calculating disk usage -- a
container with 500k files can reproduce the problem.
The executor also calculates process metrics, but the metrics are never used.
Mesos has a [posix disk
isolator|https://github.com/apache/mesos/blob/master/docs/mesos-containerizer.md]
(and XFS isolator) which provides the same functionality: it monitors disk
usage and terminates a task if it exceeds its reservation.
Thermos Observer also monitors resource usage (see AURORA-1917), so disk usage
is typically calculated 3 times -- once each by the executor, the observer, and
mesos.
This could be solved by adding [--task_process_collection_interval_secs and
--task_disk_collection_interval_secs
flags|https://github.com/apache/aurora/commit/33acb899b8cbfd9914f028524cdd9428beeb06e3]
to the executor, and if a zero interval is specified disabling resource
collection.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
