[yt-dev] Zombie jobs on eudora?

Matthew Turk matthewturk at gmail.com
Wed Jun 11 07:21:52 PDT 2014


On Tue, Jun 10, 2014 at 10:53 PM, Matthew Turk <matthewturk at gmail.com> wrote:
> On Tue, Jun 10, 2014 at 10:50 PM, Nathan Goldbaum <nathan12343 at gmail.com> wrote:
>> For the leaking YTArrays, Kacper suggested the following patch on IRC:
>>
>> http://bpaste.net/show/361120/
>>
>> This works for FLASH but seems to break field detection for enzo.
>
> I don't think this will ever be a big memory hog, but it is worth fixing.
>

I've spent a small bit of time at this again this morning, and
everything seems to come back down to the issue of having a numpy
array of grid objects.  If I switch this to a list, the reference
counting is correct again and things get deallocated properly.  I've
tried a number of ways of changing how they're allocated, but none
seem to work for getting the refcount correct.  Oddly enough, if I
track both a list *and* an array (i.e., set self._grids =
self.grids.tolist()) then the refcounting is correct.

I'm sure there's an explanation for this, but I don't know it.  It
looks to me like numpy thinks it owns the data and that it should
decrement the object refcount.

By adding this line:

self._grids = self.grids.tolist()

after the call to _populate_grid_objects() in grid_geometry_handler, I
was able to get all references tracked and removed.

>>
>>
>> On Tue, Jun 10, 2014 at 8:47 PM, Matthew Turk <matthewturk at gmail.com> wrote:
>>>
>>> Hi Nathan,
>>>
>>> I believe there are two things at work here.
>>>
>>> 1) (I do not have high confidence of this one.)  YTArrays that are
>>> referenced with a .d and turned into numpy arrays which no longer *own
>>> the data* may be retaining a reference, but that reference doesn't get
>>> freed later.  This happens often when we are doing things in the
>>> hierarchy instantiation phase.  I haven't been able to figure out
>>> which references get lost; for me, over 40 outputs, I lost 1560.  I
>>> think it's 39 YTArrays per hierarchy.  This might also be related to
>>> field detection.  I think this is not a substantial contributor.
>>> 2) For some reason, when the .grids attribute (object array) is
>>> deleted on an index, the refcounts of those grids don't decrease.  I
>>> am able to decrease their refcounts by manually setting
>>> pf.index.grids[:] = None.  This eliminated all retained grid
>>> references.
>>>
>>> So, I think the root is that at some point, because of circular
>>> references or whatever, the finalizer isn't being called on the
>>> Gridndex (or on Index itself).  This results in the reference to the
>>> grids array being kept, which then pumps up the lost object count.  I
>>> don't know why it's not getting called (it's not guaranteed to be
>>> called, in any event).
>>>
>>> I have to take care of some other things (including Brendan's note
>>> about the memory problems with particle datasets) but I am pretty sure
>>> this is the root.
>>>
>>> -Matt
>>>
>>> On Tue, Jun 10, 2014 at 10:13 PM, Matthew Turk <matthewturk at gmail.com>
>>> wrote:
>>> > Hi Nathan,
>>> >
>>> > All it requires is a call to .index; you don't need to do anything
>>> > else to get it to lose references.
>>> >
>>> > I'm still looking into it.
>>> >
>>> > -Matt
>>> >
>>> > On Tue, Jun 10, 2014 at 9:26 PM, Nathan Goldbaum <nathan12343 at gmail.com>
>>> > wrote:
>>> >>
>>> >>
>>> >>
>>> >> On Tue, Jun 10, 2014 at 10:59 AM, Matthew Turk <matthewturk at gmail.com>
>>> >> wrote:
>>> >>>
>>> >>> Do you have a reproducible script?
>>> >>
>>> >>
>>> >> This should do the trick: http://paste.yt-project.org/show/4767/
>>> >>
>>> >> (this is with an enzo dataset by the way)
>>> >>
>>> >> That script prints (on my machine):
>>> >>
>>> >> EnzoGrid    15065
>>> >> YTArray     1520
>>> >> list        704
>>> >> dict        2
>>> >> MaskedArray 1
>>> >>
>>> >> Which indicates that 15000 EnzoGrid objects and 1520 YTArray objects
>>> >> have
>>> >> leaked.
>>> >>
>>> >> The list I'm printing out at the end of the script should be the
>>> >> objects
>>> >> that leaked during the loop over the Enzo dataset.  The
>>> >> objgraph.get_leaking_objects() function returns the list of all objects
>>> >> being tracked by the garbage collector that have no references but
>>> >> still
>>> >> have nonzero refcounts.
>>> >>
>>> >> This means the "original_leaks" list isn't necessarily a list of leaky
>>> >> objects - most of the things in there are singletons that the
>>> >> interpreter
>>> >> keeps around. To create a list of leaky objects produced by iterating
>>> >> over
>>> >> the loop I take the set difference of the output of
>>> >> get_leaking_objects()
>>> >> before and after iterating over the dataset.
>>> >>
>>> >>>
>>> >>> If you make a bunch of symlinks to
>>> >>> one flash file and load them all in sequence, does that replicate the
>>> >>> behavior?
>>> >>
>>> >>
>>> >> Yes, it seems to.  Compare the output of this script:
>>> >> http://paste.yt-project.org/show/4768/
>>> >>
>>> >> Adjust the range of the for loop from 0 to 5 - creating the needed
>>> >> symlinks
>>> >> to WindTunnel/windtunnel_4lev_hdf5_plt_cnt_0040 as needed.
>>> >>
>>> >>>
>>> >>>
>>> >>> On Tue, Jun 10, 2014 at 12:57 PM, Nathan Goldbaum
>>> >>> <nathan12343 at gmail.com>
>>> >>> wrote:
>>> >>> >
>>> >>> >
>>> >>> >
>>> >>> > On Tue, Jun 10, 2014 at 10:45 AM, Matthew Turk
>>> >>> > <matthewturk at gmail.com>
>>> >>> > wrote:
>>> >>> >>
>>> >>> >> Hi Nathan,
>>> >>> >>
>>> >>> >> On Tue, Jun 10, 2014 at 12:43 PM, Nathan Goldbaum
>>> >>> >> <nathan12343 at gmail.com>
>>> >>> >> wrote:
>>> >>> >> >
>>> >>> >> >
>>> >>> >> >
>>> >>> >> > On Tue, Jun 10, 2014 at 6:09 AM, Matthew Turk
>>> >>> >> > <matthewturk at gmail.com>
>>> >>> >> > wrote:
>>> >>> >> >>
>>> >>> >> >> Hi Nathan,
>>> >>> >> >>
>>> >>> >> >> On Mon, Jun 9, 2014 at 11:02 PM, Nathan Goldbaum
>>> >>> >> >> <nathan12343 at gmail.com>
>>> >>> >> >> wrote:
>>> >>> >> >> > Hey all,
>>> >>> >> >> >
>>> >>> >> >> > I'm looking at a memory leak that Philip (cc'd) is seeing when
>>> >>> >> >> > iterating
>>> >>> >> >> > over a long list of FLASH datasets.  Just as an example of the
>>> >>> >> >> > type
>>> >>> >> >> > of
>>> >>> >> >> > behavior he is seeing - today he left his script running and
>>> >>> >> >> > ended
>>> >>> >> >> > up
>>> >>> >> >> > consuming 300 GB of RAM on a viz node.
>>> >>> >> >> >
>>> >>> >> >> > FWIW, the dataset is not particularly large - ~300 outputs and
>>> >>> >> >> > ~100
>>> >>> >> >> > MB
>>> >>> >> >> > per
>>> >>> >> >> > output. These are also FLASH cylindrical coordinate
>>> >>> >> >> > simulations -
>>> >>> >> >> > so
>>> >>> >> >> > perhaps
>>> >>> >> >> > this behavior will only occur in curvilinear geometries?
>>> >>> >> >>
>>> >>> >> >> Hm, I don't know about that.
>>> >>> >> >>
>>> >>> >> >> >
>>> >>> >> >> > I've been playing with objgraph to try to understand what's
>>> >>> >> >> > happening.
>>> >>> >> >> > Here's the script I've been using:
>>> >>> >> >> > http://paste.yt-project.org/show/4762/
>>> >>> >> >> >
>>> >>> >> >> > Here's the output after one iteration of the for loop:
>>> >>> >> >> > http://paste.yt-project.org/show/4761/
>>> >>> >> >> >
>>> >>> >> >> > It seems that for some reason a lot of data is not being
>>> >>> >> >> > garbage
>>> >>> >> >> > collected.
>>> >>> >> >> >
>>> >>> >> >> > Could there be a reference counting bug somewhere down in a
>>> >>> >> >> > cython
>>> >>> >> >> > routine?
>>> >>> >> >>
>>> >>> >> >> Based on what you're running, the only Cython routines being
>>> >>> >> >> called
>>> >>> >> >> are likely in the selection system.
>>> >>> >> >>
>>> >>> >> >> > Objgraph is unable to find backreferences to root grid tiles
>>> >>> >> >> > in
>>> >>> >> >> > the
>>> >>> >> >> > flash
>>> >>> >> >> > dataset, and all the other yt objects that I've looked at seem
>>> >>> >> >> > to
>>> >>> >> >> > have
>>> >>> >> >> > backreference graphs that terminate at a FLASHGrid object that
>>> >>> >> >> > represents a
>>> >>> >> >> > root grid tile in one of the datasets.  That's the best guess
>>> >>> >> >> > I
>>> >>> >> >> > have
>>> >>> >> >> > -
>>> >>> >> >> > but
>>> >>> >> >> > definitely nothing conclusive.  I'd appreciate any other ideas
>>> >>> >> >> > anyone
>>> >>> >> >> > else
>>> >>> >> >> > has to help debug this.
>>> >>> >> >>
>>> >>> >> >> I'm not entirely sure how to parse the output you've pasted, but
>>> >>> >> >> I
>>> >>> >> >> do
>>> >>> >> >> have a thought.  If you have a reproducible case, I can test it
>>> >>> >> >> myself.  I am wondering if this could be related to the way that
>>> >>> >> >> grid
>>> >>> >> >> masks are cached.  You should be able to test this by adding
>>> >>> >> >> this
>>> >>> >> >> line
>>> >>> >> >> to _get_selector_mask in grid_patch.py, just before "return
>>> >>> >> >> mask"
>>> >>> >> >>
>>> >>> >> >> self._last_mask = self._last_selector_id = None
>>> >>> >> >>
>>> >>> >> >> Something like this patch:
>>> >>> >> >>
>>> >>> >> >> http://paste.yt-project.org/show/4316/
>>> >>> >> >
>>> >>> >> >
>>> >>> >> > Thanks for the code!  I will look into this today.
>>> >>> >> >
>>> >>> >> > Sorry for not explaining the random terminal output I pasted from
>>> >>> >> > objgraph
>>> >>> >> > :/
>>> >>> >> >
>>> >>> >> > It's a list of objects created after yt operates on one dataset
>>> >>> >> > and
>>> >>> >> > after
>>> >>> >> > the garbage collector is explicitly called. Each iteration of the
>>> >>> >> > loop
>>> >>> >> > sees
>>> >>> >> > the creation of objects representing the FLASH grids, hierarchy,
>>> >>> >> > and
>>> >>> >> > associated metadata.  With enough iterations this overhead from
>>> >>> >> > previous
>>> >>> >> > loop iterations begins to dominate the total memory budget.
>>> >>> >>
>>> >>> >> The code snippet I sent might help reduce it, but I think it speaks
>>> >>> >> to
>>> >>> >> a deeper problem in that somehow the FLASH stuff isn't being GC'd
>>> >>> >> anywhere.  It really ought to be.
>>> >>> >>
>>> >>> >> Can you try also doing:
>>> >>> >>
>>> >>> >> yt.frontends.flash.FLASHDataset._skip_cache = True
>>> >>> >
>>> >>> >
>>> >>> > No effect, unfortunately.
>>> >>> >
>>> >>> >>
>>> >>> >> and seeing if that helps?
>>> >>> >>
>>> >>> >> >
>>> >>> >> >>
>>> >>> >> >>
>>> >>> >> >>
>>> >>> >> >> -Matt
>>> >>> >> >>
>>> >>> >> >> >
>>> >>> >> >> > Thanks for your help in debugging this!
>>> >>> >> >> >
>>> >>> >> >> > -Nathan
>>> >>> >> >> >
>>> >>> >> >> _______________________________________________
>>> >>> >> >> yt-dev mailing list
>>> >>> >> >> yt-dev at lists.spacepope.org
>>> >>> >> >> http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
>>> >>> >> >
>>> >>> >> >
>>> >>> >> >
>>> >>> >> > _______________________________________________
>>> >>> >> > yt-dev mailing list
>>> >>> >> > yt-dev at lists.spacepope.org
>>> >>> >> > http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
>>> >>> >> >
>>> >>> >> _______________________________________________
>>> >>> >> yt-dev mailing list
>>> >>> >> yt-dev at lists.spacepope.org
>>> >>> >> http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
>>> >>> >
>>> >>> >
>>> >>> >
>>> >>> > _______________________________________________
>>> >>> > yt-dev mailing list
>>> >>> > yt-dev at lists.spacepope.org
>>> >>> > http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
>>> >>> >
>>> >>> _______________________________________________
>>> >>> yt-dev mailing list
>>> >>> yt-dev at lists.spacepope.org
>>> >>> http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
>>> >>
>>> >>
>>> >>
>>> >> _______________________________________________
>>> >> yt-dev mailing list
>>> >> yt-dev at lists.spacepope.org
>>> >> http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
>>> >>
>>> _______________________________________________
>>> yt-dev mailing list
>>> yt-dev at lists.spacepope.org
>>> http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
>>
>>
>>
>> _______________________________________________
>> yt-dev mailing list
>> yt-dev at lists.spacepope.org
>> http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org
>>



More information about the yt-dev mailing list