mesos-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Qian Zhang" <zhang...@cn.ibm.com>
Subject Re: Review Request 41305: MESOS-1718: move getExecutorInfo from slave to master
Date Tue, 15 Dec 2015 03:33:10 GMT


> On Dec. 14, 2015, 2:31 p.m., Qian Zhang wrote:
> > src/master/master.cpp, line 2830
> > <https://reviews.apache.org/r/41305/diff/1/?file=1161526#file1161526line2830>
> >
> >     So if task's resources has no cpu, then we will not add cpu to executor's resources,
right? But I think we should always add DEFAULT_EXECUTOR_CPUS to executor's resources in this
case.
> 
> Klaus Ma wrote:
>     The purpose of this code is not to overcommit resources; for example, only CPU or
memory in cluster, master can not assign more resources to it. It may exhaust slave's resource
if lots of executor in slave.

So that means, it is possible to launch a command executor with no cpu and no memory? But
when a command executor is running in an agent, it will indeed consume cpu and memory in that
agent, but those cpu and memory will not be tracked by allocator for that agent with the code
you wrote here, is that a resource leak in that agent? I think this will actually cause the
resource overcommit in the agent.


> On Dec. 14, 2015, 2:31 p.m., Qian Zhang wrote:
> > src/master/master.cpp, line 3556
> > <https://reviews.apache.org/r/41305/diff/1/?file=1161526#file1161526line3556>
> >
> >     Just wondering why we want to do `task.resources() - executor.resources())`
here? I think in `validateResourceUsage()`, we will validate whether the task's resources
+ executor's resources can be contained by the offered resources, if here we substract executor's
resources from task's resources, then we will only validate whether task's resources can be
contained by the offered resources in `validateResourceUsage()` which seems not correct to
me.
> 
> Klaus Ma wrote:
>     I do not want to overcommit the command line tasks; for example, if there're only
2 CPU, how may tasks (~1 CPU per task) should framework run?

Understand your concern. But what about framework launches a small task (e.g., 0.1 cpu and
32MB memory)? So if we substract executor's resources from task's resources here, that means
we will launch a task with empty resources?


- Qian


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/41305/#review110166
-----------------------------------------------------------


On Dec. 12, 2015, 5:54 p.m., Klaus Ma wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/41305/
> -----------------------------------------------------------
> 
> (Updated Dec. 12, 2015, 5:54 p.m.)
> 
> 
> Review request for mesos, Ben Mahler, Guangya Liu, Artem Harutyunyan, Ian Downes, Joris
Van Remoortere, Joseph Wu, and Vinod Kone.
> 
> 
> Bugs: MESOS-1718
>     https://issues.apache.org/jira/browse/MESOS-1718
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> MESOS-1718: move getExecutorInfo from slave to master
> 
> 
> Diffs
> -----
> 
>   src/master/constants.hpp cc38dfcf57ecbc8555379084acbf2bfa4b3fc759 
>   src/master/master.hpp 9aa548aa6e159046c94e4ec96f631ea8b3bfd5d8 
>   src/master/master.cpp 0d1482279c68f2a4a27dabaf28774769a5d515c4 
>   src/slave/constants.hpp bcbb1401aa8f9f04c4f9256bb4f560e18d8994e0 
>   src/slave/slave.hpp b7586ce42bfac9d9885a3eb8d82deb94680c236c 
>   src/slave/slave.cpp 9bd86e122c070cd072a54d4de8097f844bd95bb0 
> 
> Diff: https://reviews.apache.org/r/41305/diff/
> 
> 
> Testing
> -------
> 
> 
> Thanks,
> 
> Klaus Ma
> 
>


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