> On July 17, 2013, 2:58 a.m., Ben Mahler wrote:
> > src/sched/sched.cpp, line 78
> > <https://reviews.apache.org/r/12603/diff/2/?file=322317#file322317line78>
> >
> >     CACHED_UPDATES_CAPACITY?
> >     
> >     Also, maybe add a note that this only works for smaller clusters. If 
> > there are more than 100 offers in 10 seconds, this will not eliminate any 
> > duplicates and will only be adding unnecessary overhead, right?
> >     
> >     Caching O(20-30) seconds of updates seems nicer, but we'd have to be 
> > careful of the memory / performance overhead. Thoughts?
> 
> Vinod Kone wrote:
>     Changed it to "MAX_ACKED_UPDATES". I think thats more descriptive.
>     
>     Regarding the number, it was definitely pulled out of thin air. Thanks to 
> your nudge, here are my thoughts.
>     
>     As you alluded, the critical number here is the "number of unique status 
> updates" / "status update retry interval" generated by the cluster. 
> Currently, the retry interval is set to 10s. Note that if the rate of status 
> unique updates is high, the changes in this patch are irrelevant. The 
> scheduler would still get backed up processing the updates.
>     
>     Now regarding MAX_ACKED_UPDATES value, the higher the number the more 
> likely the driver can bypass the scheduler for duplicates. As far as I can 
> tell, the only concern would be the memory consumed. Back of envelope math 
> suggests it is roughly, 2 * 16 bytes (UUID size) *  MAX_ACKED_UPDATES. For 
> MAX_ACKED_UPDATES=1000, this value is ~32KB whereas for 10,000 it is ~320KB. 
> IMO, ~320KB overhead for the driver is reasonable, so I will bump this number 
> to 10,000.
>     
>     Thoughts?

That's a much better description that what I said! Care to add a comment 
reflecting this (the part about the status update rate) in the code for 
posterity?

SGTM, I'm ok caching O(1000-10,000) updates. This is useful for calculating the 
memory overhead: 
http://www.boost.org/doc/libs/1_54_0/doc/html/intrusive/unordered_set_unordered_multiset.html#intrusive.unordered_set_unordered_multiset.unordered_set_unordered_multiset_performance


- Ben


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/12603/#review23230
-----------------------------------------------------------


On July 17, 2013, 6:32 a.m., Vinod Kone wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/12603/
> -----------------------------------------------------------
> 
> (Updated July 17, 2013, 6:32 a.m.)
> 
> 
> Review request for mesos, Benjamin Hindman and Ben Mahler.
> 
> 
> Bugs: MESOS-551
>     https://issues.apache.org/jira/browse/MESOS-551
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> See summary.
> 
> 
> Diffs
> -----
> 
>   src/sched/sched.cpp 7ea82e547c612159c9fa24fb6d62e3d2b5f11982 
>   src/tests/status_update_manager_tests.cpp 
> 42395324dfe49659bee2229c6573ffef0874d923 
> 
> Diff: https://reviews.apache.org/r/12603/diff/
> 
> 
> Testing
> -------
> 
> make check (OSX and Linux)
> 
> 
> Thanks,
> 
> Vinod Kone
> 
>

Reply via email to