The coercion model is keeping references to tracebacks which causes memory leaks.

[Maarten Derickx]

When I was doing some computations on modular forms, I noticed that the ModularSymbols_clear_cache() was not doing what it claims to do.

sage: import gc  
sage: m=ModularSymbols(Gamma1(29),sign=1)
sage: m=[]
sage: ModularSymbols_clear_cache()
sage: gc.collect()
57
sage: [x for x in gc.get_objects() if isinstance(x,sage.modular.modsym.ambient.ModularSymbolsAmbient_wtk_g1)]
[Modular Symbols space of dimension 49 for Gamma_1(29) of weight 2 with sign 1 and over Rational Field]
sage: 

So even after garbage collection the modular symbols element is still in memory.

When testing please also make sure that the last stackframe from #10570 is also gone.

Apply 10548-coerce-traceback.2.patch​ and trac_10548-coerce-traceback-doctest.v2.patch​

[Maarten Derickx]

Ps. you might want to use smaller numbers then 97 to test because ModularSymbols?(Gamma1(97),sign=1) takes a while. It behaves the same, but with smaller memory usage for smaller numbers.

[Maarten Derickx]

According to heapy the aditional memory comes from;

dict of sage.modular.modsym.manin_symbols.ManinSymbol?

mderickx@mod:~$ sage
----------------------------------------------------------------------
| Sage Version 4.6, Release Date: 2010-10-30                         |
| Type notebook() for the GUI, and license() for information.        |
----------------------------------------------------------------------
sage: import gc
sage: from guppy import hpy; hp=hpy()
sage: M=ModularSymbols(Gamma1(97),sign=1)
sage: get_memory_usage()
848.1953125
sage: M=[]
sage: get_memory_usage()
848.1953125
sage: ModularSymbols_clear_cache()
sage: get_memory_usage()
848.1953125
sage: hp.heap()
Partition of a set of 445281 objects. Total size = 68518056 bytes.
 Index  Count   %     Size   % Cumulative  % Kind (class / dict of class)
     0 169534  38 28541776  42  28541776  42 str
     1 125494  28 10674696  16  39216472  57 tuple
     2   1624   0  3964480   6  43180952  63 dict of module
     3  25817   6  3098040   5  46278992  68 types.CodeType
     4  24634   6  2956080   4  49235072  72 function
     5   2798   1  2949584   4  52184656  76 dict (no owner)
     6   2660   1  2727392   4  54912048  80 dict of type
     7   9408   2  2634240   4  57546288  84 dict of sage.modular.modsym.manin_symbols.ManinSymbol
     8   2660   1  2388040   3  59934328  87 type
     9   4293   1  1010536   1  60944864  89 list
<895 more rows. Type e.g. '_.more' to view.>
sage: 

[Maarten Derickx]

It's not an error of the ModularSymbols_clear_cache() function

sage: ModularSymbols(Gamma1(29),sign=1,use_cache=False)
Modular Symbols space of dimension 49 for Gamma_1(29) of weight 2 with sign 1 and over Rational Field
sage: None
sage: gc.collect()
54
sage: [x for x in gc.get_objects() if isinstance(x,sage.modular.modsym.ambient.ModularSymbolsAmbient_wtk_g1)]
[Modular Symbols space of dimension 49 for Gamma_1(29) of weight 2 with sign 1 and over Rational Field]

[Maarten Derickx]

See the attached chain.png picture created with objgraph. There is a reference chain to the homset module. And the homeset module will not get destroyed so we have to find out where this chain comes from.

[David Loeffler]

Might this be an instance of #715? Modular symbols spaces are Parents, so various maps in and out of them are probably getting cached by the coercion model machinery.

[Maarten Derickx]

Sorry I failed in adding the mentioned chain.png before. It shows a reference chain explaining why ManinSymbol?(37,35) is still in the memory. It seems that it has at least to do something with code in the homset module. I don't know enough details about the coercion and category framework and their interaction to say that it has something to do with 715. But the description of 715 really looks like it could be the cause.

If it is coercion, it might also be related to: #10570

[David Loeffler]

Having looked at #10570, I don't think that's the reason. #715 seems a much more likely culprit.

[Maarten Derickx]

Here is some code run on sage 4.6.1 showing that it is really the coercion framework.

sage: import gc  
sage: import objgraph
sage: import inspect
sage: m=ModularSymbols(Gamma1(29),use_cache=False)
sage: m=[]
sage: ModularSymbols_clear_cache()
sage: gc.collect()
54
sage: a=[x for x in gc.get_objects() if isinstance(x,sage.modular.modsym.ambient.ModularSymbolsAmbient_wtk_g1)]
sage: l=objgraph.find_backref_chain(a[0],inspect.ismodule,max_depth=15,extra_ignore=[id(a)])
sage: map(type,l)
[<type 'module'>, <type 'dict'>, <type 'sage.structure.coerce.CoercionModel_cache_maps'>, <type 'list'>, <type 'tuple'>, <type 'traceback'>, <type 'frame'>, <type 'frame'>, <type 'frame'>, <type 'frame'>, <class 'sage.modular.modsym.ambient.ModularSymbolsAmbient_wtk_g1_with_category'>]

[Maarten Derickx]

btw it goes wrong in the same way on 4.7.alpha1

[Robert Bradshaw]

The attached patch fixes the issue described in the title, but does not fix #715.

[Maarten Derickx]

rebased against 4.7.alpha1

[Maarten Derickx]

Looks good to me.

[Maarten Derickx]

Robert, (or someone else), can you review the doctest I added?

[Robert Bradshaw]

The doctest looks good to me.

[Keshav Kini]

new doctest

[Keshav Kini]

I folded in the doctest from the current patch at #10570, also by Maarten.