> On June 20, 2013, 10:20 p.m., Ben Mahler wrote:
> > src/slave/constants.cpp, line 32
> > <https://reviews.apache.org/r/11100/diff/2/?file=309822#file309822line32>
> >
> >     Ok, so 10 per second maximum.
> >     
> >     Let's say we run 4 tasks per minute, that have 50 files / directories 
> > to remove. That's 200/minute -> 3.33 files / minute. Seems like we might 
> > need to implement a strategy for handling a growing queue size of removals 
> > (by having a dynamic interval), can you add a TODO signaling this potential 
> > issue?
> 
> Vinod Kone wrote:
>     Agreed on the need to be smarter about this. But not sure a growing queue 
> size implies a need to do it faster. What we are trying to do here is to 
> maintain a constant disk i/o. Just because there are more files to remove, 
> doesn't mean we can do it faster. It could increase the disk i/o to the point 
> of affecting the slave. I think we need smartness in figuring our the right 
> interval. Does that make sense?
>     
>     
>

I think number of files to delete is a weak proxy for disk i/o. Is this to 
prevent a storm of random iops introducing latency to other processes?

Wouldn't this be better phrased as i/o prioritization rather than rate limiting 
the deletion: opportunistically process the file deletion queue as fast as 
possible but yield to others when they have any i/o. Include baseline i/o to 
ensure it's sufficiently serviced.

Furthermore, couldn't this prioritization be pushed to the kernel i/o scheduler 
itself? Have either ionice(1) or cgroups/blkio been considered?


- Ian


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/11100/#review22201
-----------------------------------------------------------


On June 21, 2013, 12:44 a.m., Vinod Kone wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/11100/
> -----------------------------------------------------------
> 
> (Updated June 21, 2013, 12:44 a.m.)
> 
> 
> Review request for mesos, Benjamin Hindman and Ben Mahler.
> 
> 
> Description
> -------
> 
> The purpose of the incinerator is to move items to trash and empty trash in 
> rate-limited manner. It is designed to work across slave restarts.
> 
> 
> Diffs
> -----
> 
>   src/Makefile.am 933743593d8de2fcc412d2f771a4cbc1e6fd2884 
>   src/slave/constants.hpp 901fdf220a902de9241511393530eb19fdfc3244 
>   src/slave/constants.cpp e8d16ca3307249a8b49720eaf8dcb0e7555fca7a 
>   src/slave/incinerator.hpp PRE-CREATION 
>   src/slave/incinerator.cpp PRE-CREATION 
>   src/tests/gc_tests.cpp 174580086370f61be369f7a212d62531e87e6f01 
> 
> Diff: https://reviews.apache.org/r/11100/diff/
> 
> 
> Testing
> -------
> 
> make check
> 
> 
> Thanks,
> 
> Vinod Kone
> 
>

Reply via email to