<div dir="ltr">Could this issue be related?<div><br></div><div><a href="https://github.com/numpy/numpy/issues/1601">https://github.com/numpy/numpy/issues/1601</a><br></div><div><br></div><div>Can you elaborate a bit more about why we're using an object array in the first place? If switching to using a list solves these issues perhaps that is the way to go.</div>
</div><div class="gmail_extra"><br><br><div class="gmail_quote">On Wed, Jun 11, 2014 at 7:25 AM, Matthew Turk <span dir="ltr"><<a href="mailto:matthewturk@gmail.com" target="_blank">matthewturk@gmail.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">I should also note, at some point in the future I want to get rid of<br>
the object arrays for grids, but that timescale is longer. Using<br>
John's grid tree is a much better approach.<br>
<div class="HOEnZb"><div class="h5"><br>
On Wed, Jun 11, 2014 at 9:21 AM, Matthew Turk <<a href="mailto:matthewturk@gmail.com">matthewturk@gmail.com</a>> wrote:<br>
> On Tue, Jun 10, 2014 at 10:53 PM, Matthew Turk <<a href="mailto:matthewturk@gmail.com">matthewturk@gmail.com</a>> wrote:<br>
>> On Tue, Jun 10, 2014 at 10:50 PM, Nathan Goldbaum <<a href="mailto:nathan12343@gmail.com">nathan12343@gmail.com</a>> wrote:<br>
>>> For the leaking YTArrays, Kacper suggested the following patch on IRC:<br>
>>><br>
>>> <a href="http://bpaste.net/show/361120/" target="_blank">http://bpaste.net/show/361120/</a><br>
>>><br>
>>> This works for FLASH but seems to break field detection for enzo.<br>
>><br>
>> I don't think this will ever be a big memory hog, but it is worth fixing.<br>
>><br>
><br>
> I've spent a small bit of time at this again this morning, and<br>
> everything seems to come back down to the issue of having a numpy<br>
> array of grid objects. If I switch this to a list, the reference<br>
> counting is correct again and things get deallocated properly. I've<br>
> tried a number of ways of changing how they're allocated, but none<br>
> seem to work for getting the refcount correct. Oddly enough, if I<br>
> track both a list *and* an array (i.e., set self._grids =<br>
> self.grids.tolist()) then the refcounting is correct.<br>
><br>
> I'm sure there's an explanation for this, but I don't know it. It<br>
> looks to me like numpy thinks it owns the data and that it should<br>
> decrement the object refcount.<br>
><br>
> By adding this line:<br>
><br>
> self._grids = self.grids.tolist()<br>
><br>
> after the call to _populate_grid_objects() in grid_geometry_handler, I<br>
> was able to get all references tracked and removed.<br>
><br>
>>><br>
>>><br>
>>> On Tue, Jun 10, 2014 at 8:47 PM, Matthew Turk <<a href="mailto:matthewturk@gmail.com">matthewturk@gmail.com</a>> wrote:<br>
>>>><br>
>>>> Hi Nathan,<br>
>>>><br>
>>>> I believe there are two things at work here.<br>
>>>><br>
>>>> 1) (I do not have high confidence of this one.) YTArrays that are<br>
>>>> referenced with a .d and turned into numpy arrays which no longer *own<br>
>>>> the data* may be retaining a reference, but that reference doesn't get<br>
>>>> freed later. This happens often when we are doing things in the<br>
>>>> hierarchy instantiation phase. I haven't been able to figure out<br>
>>>> which references get lost; for me, over 40 outputs, I lost 1560. I<br>
>>>> think it's 39 YTArrays per hierarchy. This might also be related to<br>
>>>> field detection. I think this is not a substantial contributor.<br>
>>>> 2) For some reason, when the .grids attribute (object array) is<br>
>>>> deleted on an index, the refcounts of those grids don't decrease. I<br>
>>>> am able to decrease their refcounts by manually setting<br>
>>>> pf.index.grids[:] = None. This eliminated all retained grid<br>
>>>> references.<br>
>>>><br>
>>>> So, I think the root is that at some point, because of circular<br>
>>>> references or whatever, the finalizer isn't being called on the<br>
>>>> Gridndex (or on Index itself). This results in the reference to the<br>
>>>> grids array being kept, which then pumps up the lost object count. I<br>
>>>> don't know why it's not getting called (it's not guaranteed to be<br>
>>>> called, in any event).<br>
>>>><br>
>>>> I have to take care of some other things (including Brendan's note<br>
>>>> about the memory problems with particle datasets) but I am pretty sure<br>
>>>> this is the root.<br>
>>>><br>
>>>> -Matt<br>
>>>><br>
>>>> On Tue, Jun 10, 2014 at 10:13 PM, Matthew Turk <<a href="mailto:matthewturk@gmail.com">matthewturk@gmail.com</a>><br>
>>>> wrote:<br>
>>>> > Hi Nathan,<br>
>>>> ><br>
>>>> > All it requires is a call to .index; you don't need to do anything<br>
>>>> > else to get it to lose references.<br>
>>>> ><br>
>>>> > I'm still looking into it.<br>
>>>> ><br>
>>>> > -Matt<br>
>>>> ><br>
>>>> > On Tue, Jun 10, 2014 at 9:26 PM, Nathan Goldbaum <<a href="mailto:nathan12343@gmail.com">nathan12343@gmail.com</a>><br>
>>>> > wrote:<br>
>>>> >><br>
>>>> >><br>
>>>> >><br>
>>>> >> On Tue, Jun 10, 2014 at 10:59 AM, Matthew Turk <<a href="mailto:matthewturk@gmail.com">matthewturk@gmail.com</a>><br>
>>>> >> wrote:<br>
>>>> >>><br>
>>>> >>> Do you have a reproducible script?<br>
>>>> >><br>
>>>> >><br>
>>>> >> This should do the trick: <a href="http://paste.yt-project.org/show/4767/" target="_blank">http://paste.yt-project.org/show/4767/</a><br>
>>>> >><br>
>>>> >> (this is with an enzo dataset by the way)<br>
>>>> >><br>
>>>> >> That script prints (on my machine):<br>
>>>> >><br>
>>>> >> EnzoGrid 15065<br>
>>>> >> YTArray 1520<br>
>>>> >> list 704<br>
>>>> >> dict 2<br>
>>>> >> MaskedArray 1<br>
>>>> >><br>
>>>> >> Which indicates that 15000 EnzoGrid objects and 1520 YTArray objects<br>
>>>> >> have<br>
>>>> >> leaked.<br>
>>>> >><br>
>>>> >> The list I'm printing out at the end of the script should be the<br>
>>>> >> objects<br>
>>>> >> that leaked during the loop over the Enzo dataset. The<br>
>>>> >> objgraph.get_leaking_objects() function returns the list of all objects<br>
>>>> >> being tracked by the garbage collector that have no references but<br>
>>>> >> still<br>
>>>> >> have nonzero refcounts.<br>
>>>> >><br>
>>>> >> This means the "original_leaks" list isn't necessarily a list of leaky<br>
>>>> >> objects - most of the things in there are singletons that the<br>
>>>> >> interpreter<br>
>>>> >> keeps around. To create a list of leaky objects produced by iterating<br>
>>>> >> over<br>
>>>> >> the loop I take the set difference of the output of<br>
>>>> >> get_leaking_objects()<br>
>>>> >> before and after iterating over the dataset.<br>
>>>> >><br>
>>>> >>><br>
>>>> >>> If you make a bunch of symlinks to<br>
>>>> >>> one flash file and load them all in sequence, does that replicate the<br>
>>>> >>> behavior?<br>
>>>> >><br>
>>>> >><br>
>>>> >> Yes, it seems to. Compare the output of this script:<br>
>>>> >> <a href="http://paste.yt-project.org/show/4768/" target="_blank">http://paste.yt-project.org/show/4768/</a><br>
>>>> >><br>
>>>> >> Adjust the range of the for loop from 0 to 5 - creating the needed<br>
>>>> >> symlinks<br>
>>>> >> to WindTunnel/windtunnel_4lev_hdf5_plt_cnt_0040 as needed.<br>
>>>> >><br>
>>>> >>><br>
>>>> >>><br>
>>>> >>> On Tue, Jun 10, 2014 at 12:57 PM, Nathan Goldbaum<br>
>>>> >>> <<a href="mailto:nathan12343@gmail.com">nathan12343@gmail.com</a>><br>
>>>> >>> wrote:<br>
>>>> >>> ><br>
>>>> >>> ><br>
>>>> >>> ><br>
>>>> >>> > On Tue, Jun 10, 2014 at 10:45 AM, Matthew Turk<br>
>>>> >>> > <<a href="mailto:matthewturk@gmail.com">matthewturk@gmail.com</a>><br>
>>>> >>> > wrote:<br>
>>>> >>> >><br>
>>>> >>> >> Hi Nathan,<br>
>>>> >>> >><br>
>>>> >>> >> On Tue, Jun 10, 2014 at 12:43 PM, Nathan Goldbaum<br>
>>>> >>> >> <<a href="mailto:nathan12343@gmail.com">nathan12343@gmail.com</a>><br>
>>>> >>> >> wrote:<br>
>>>> >>> >> ><br>
>>>> >>> >> ><br>
>>>> >>> >> ><br>
>>>> >>> >> > On Tue, Jun 10, 2014 at 6:09 AM, Matthew Turk<br>
>>>> >>> >> > <<a href="mailto:matthewturk@gmail.com">matthewturk@gmail.com</a>><br>
>>>> >>> >> > wrote:<br>
>>>> >>> >> >><br>
>>>> >>> >> >> Hi Nathan,<br>
>>>> >>> >> >><br>
>>>> >>> >> >> On Mon, Jun 9, 2014 at 11:02 PM, Nathan Goldbaum<br>
>>>> >>> >> >> <<a href="mailto:nathan12343@gmail.com">nathan12343@gmail.com</a>><br>
>>>> >>> >> >> wrote:<br>
>>>> >>> >> >> > Hey all,<br>
>>>> >>> >> >> ><br>
>>>> >>> >> >> > I'm looking at a memory leak that Philip (cc'd) is seeing when<br>
>>>> >>> >> >> > iterating<br>
>>>> >>> >> >> > over a long list of FLASH datasets. Just as an example of the<br>
>>>> >>> >> >> > type<br>
>>>> >>> >> >> > of<br>
>>>> >>> >> >> > behavior he is seeing - today he left his script running and<br>
>>>> >>> >> >> > ended<br>
>>>> >>> >> >> > up<br>
>>>> >>> >> >> > consuming 300 GB of RAM on a viz node.<br>
>>>> >>> >> >> ><br>
>>>> >>> >> >> > FWIW, the dataset is not particularly large - ~300 outputs and<br>
>>>> >>> >> >> > ~100<br>
>>>> >>> >> >> > MB<br>
>>>> >>> >> >> > per<br>
>>>> >>> >> >> > output. These are also FLASH cylindrical coordinate<br>
>>>> >>> >> >> > simulations -<br>
>>>> >>> >> >> > so<br>
>>>> >>> >> >> > perhaps<br>
>>>> >>> >> >> > this behavior will only occur in curvilinear geometries?<br>
>>>> >>> >> >><br>
>>>> >>> >> >> Hm, I don't know about that.<br>
>>>> >>> >> >><br>
>>>> >>> >> >> ><br>
>>>> >>> >> >> > I've been playing with objgraph to try to understand what's<br>
>>>> >>> >> >> > happening.<br>
>>>> >>> >> >> > Here's the script I've been using:<br>
>>>> >>> >> >> > <a href="http://paste.yt-project.org/show/4762/" target="_blank">http://paste.yt-project.org/show/4762/</a><br>
>>>> >>> >> >> ><br>
>>>> >>> >> >> > Here's the output after one iteration of the for loop:<br>
>>>> >>> >> >> > <a href="http://paste.yt-project.org/show/4761/" target="_blank">http://paste.yt-project.org/show/4761/</a><br>
>>>> >>> >> >> ><br>
>>>> >>> >> >> > It seems that for some reason a lot of data is not being<br>
>>>> >>> >> >> > garbage<br>
>>>> >>> >> >> > collected.<br>
>>>> >>> >> >> ><br>
>>>> >>> >> >> > Could there be a reference counting bug somewhere down in a<br>
>>>> >>> >> >> > cython<br>
>>>> >>> >> >> > routine?<br>
>>>> >>> >> >><br>
>>>> >>> >> >> Based on what you're running, the only Cython routines being<br>
>>>> >>> >> >> called<br>
>>>> >>> >> >> are likely in the selection system.<br>
>>>> >>> >> >><br>
>>>> >>> >> >> > Objgraph is unable to find backreferences to root grid tiles<br>
>>>> >>> >> >> > in<br>
>>>> >>> >> >> > the<br>
>>>> >>> >> >> > flash<br>
>>>> >>> >> >> > dataset, and all the other yt objects that I've looked at seem<br>
>>>> >>> >> >> > to<br>
>>>> >>> >> >> > have<br>
>>>> >>> >> >> > backreference graphs that terminate at a FLASHGrid object that<br>
>>>> >>> >> >> > represents a<br>
>>>> >>> >> >> > root grid tile in one of the datasets. That's the best guess<br>
>>>> >>> >> >> > I<br>
>>>> >>> >> >> > have<br>
>>>> >>> >> >> > -<br>
>>>> >>> >> >> > but<br>
>>>> >>> >> >> > definitely nothing conclusive. I'd appreciate any other ideas<br>
>>>> >>> >> >> > anyone<br>
>>>> >>> >> >> > else<br>
>>>> >>> >> >> > has to help debug this.<br>
>>>> >>> >> >><br>
>>>> >>> >> >> I'm not entirely sure how to parse the output you've pasted, but<br>
>>>> >>> >> >> I<br>
>>>> >>> >> >> do<br>
>>>> >>> >> >> have a thought. If you have a reproducible case, I can test it<br>
>>>> >>> >> >> myself. I am wondering if this could be related to the way that<br>
>>>> >>> >> >> grid<br>
>>>> >>> >> >> masks are cached. You should be able to test this by adding<br>
>>>> >>> >> >> this<br>
>>>> >>> >> >> line<br>
>>>> >>> >> >> to _get_selector_mask in grid_patch.py, just before "return<br>
>>>> >>> >> >> mask"<br>
>>>> >>> >> >><br>
>>>> >>> >> >> self._last_mask = self._last_selector_id = None<br>
>>>> >>> >> >><br>
>>>> >>> >> >> Something like this patch:<br>
>>>> >>> >> >><br>
>>>> >>> >> >> <a href="http://paste.yt-project.org/show/4316/" target="_blank">http://paste.yt-project.org/show/4316/</a><br>
>>>> >>> >> ><br>
>>>> >>> >> ><br>
>>>> >>> >> > Thanks for the code! I will look into this today.<br>
>>>> >>> >> ><br>
>>>> >>> >> > Sorry for not explaining the random terminal output I pasted from<br>
>>>> >>> >> > objgraph<br>
>>>> >>> >> > :/<br>
>>>> >>> >> ><br>
>>>> >>> >> > It's a list of objects created after yt operates on one dataset<br>
>>>> >>> >> > and<br>
>>>> >>> >> > after<br>
>>>> >>> >> > the garbage collector is explicitly called. Each iteration of the<br>
>>>> >>> >> > loop<br>
>>>> >>> >> > sees<br>
>>>> >>> >> > the creation of objects representing the FLASH grids, hierarchy,<br>
>>>> >>> >> > and<br>
>>>> >>> >> > associated metadata. With enough iterations this overhead from<br>
>>>> >>> >> > previous<br>
>>>> >>> >> > loop iterations begins to dominate the total memory budget.<br>
>>>> >>> >><br>
>>>> >>> >> The code snippet I sent might help reduce it, but I think it speaks<br>
>>>> >>> >> to<br>
>>>> >>> >> a deeper problem in that somehow the FLASH stuff isn't being GC'd<br>
>>>> >>> >> anywhere. It really ought to be.<br>
>>>> >>> >><br>
>>>> >>> >> Can you try also doing:<br>
>>>> >>> >><br>
>>>> >>> >> yt.frontends.flash.FLASHDataset._skip_cache = True<br>
>>>> >>> ><br>
>>>> >>> ><br>
>>>> >>> > No effect, unfortunately.<br>
>>>> >>> ><br>
>>>> >>> >><br>
>>>> >>> >> and seeing if that helps?<br>
>>>> >>> >><br>
>>>> >>> >> ><br>
>>>> >>> >> >><br>
>>>> >>> >> >><br>
>>>> >>> >> >><br>
>>>> >>> >> >> -Matt<br>
>>>> >>> >> >><br>
>>>> >>> >> >> ><br>
>>>> >>> >> >> > Thanks for your help in debugging this!<br>
>>>> >>> >> >> ><br>
>>>> >>> >> >> > -Nathan<br>
>>>> >>> >> >> ><br>
>>>> >>> >> >> _______________________________________________<br>
>>>> >>> >> >> yt-dev mailing list<br>
>>>> >>> >> >> <a href="mailto:yt-dev@lists.spacepope.org">yt-dev@lists.spacepope.org</a><br>
>>>> >>> >> >> <a href="http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org" target="_blank">http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org</a><br>
>>>> >>> >> ><br>
>>>> >>> >> ><br>
>>>> >>> >> ><br>
>>>> >>> >> > _______________________________________________<br>
>>>> >>> >> > yt-dev mailing list<br>
>>>> >>> >> > <a href="mailto:yt-dev@lists.spacepope.org">yt-dev@lists.spacepope.org</a><br>
>>>> >>> >> > <a href="http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org" target="_blank">http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org</a><br>
>>>> >>> >> ><br>
>>>> >>> >> _______________________________________________<br>
>>>> >>> >> yt-dev mailing list<br>
>>>> >>> >> <a href="mailto:yt-dev@lists.spacepope.org">yt-dev@lists.spacepope.org</a><br>
>>>> >>> >> <a href="http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org" target="_blank">http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org</a><br>
>>>> >>> ><br>
>>>> >>> ><br>
>>>> >>> ><br>
>>>> >>> > _______________________________________________<br>
>>>> >>> > yt-dev mailing list<br>
>>>> >>> > <a href="mailto:yt-dev@lists.spacepope.org">yt-dev@lists.spacepope.org</a><br>
>>>> >>> > <a href="http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org" target="_blank">http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org</a><br>
>>>> >>> ><br>
>>>> >>> _______________________________________________<br>
>>>> >>> yt-dev mailing list<br>
>>>> >>> <a href="mailto:yt-dev@lists.spacepope.org">yt-dev@lists.spacepope.org</a><br>
>>>> >>> <a href="http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org" target="_blank">http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org</a><br>
>>>> >><br>
>>>> >><br>
>>>> >><br>
>>>> >> _______________________________________________<br>
>>>> >> yt-dev mailing list<br>
>>>> >> <a href="mailto:yt-dev@lists.spacepope.org">yt-dev@lists.spacepope.org</a><br>
>>>> >> <a href="http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org" target="_blank">http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org</a><br>
>>>> >><br>
>>>> _______________________________________________<br>
>>>> yt-dev mailing list<br>
>>>> <a href="mailto:yt-dev@lists.spacepope.org">yt-dev@lists.spacepope.org</a><br>
>>>> <a href="http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org" target="_blank">http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org</a><br>
>>><br>
>>><br>
>>><br>
>>> _______________________________________________<br>
>>> yt-dev mailing list<br>
>>> <a href="mailto:yt-dev@lists.spacepope.org">yt-dev@lists.spacepope.org</a><br>
>>> <a href="http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org" target="_blank">http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org</a><br>
>>><br>
_______________________________________________<br>
yt-dev mailing list<br>
<a href="mailto:yt-dev@lists.spacepope.org">yt-dev@lists.spacepope.org</a><br>
<a href="http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org" target="_blank">http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org</a><br>
</div></div></blockquote></div><br></div>