Opened 21 months ago
Last modified 18 months ago
#23138 needs_work enhancement
Cache assumptions and only send to Maxima when needed
Reported by:  schymans  Owned by:  

Priority:  major  Milestone:  sage8.0 
Component:  performance  Keywords:  Maxima, symbolics, assume 
Cc:  rws, egourgoulhon  Merged in:  
Authors:  Ralf Stephan  Reviewers:  
Report Upstream:  N/A  Work issues:  
Branch:  u/rws/cache_assumptions_and_only_send_to_maxima_when_needed (Commits)  Commit:  10c31a1971edb7c32fd833d9a497a7ade393471f 
Dependencies:  #23325  Stopgaps: 
Description (last modified by )
As described in https://groups.google.com/forum/#!topic/sagedevel/jN6inWPyElM, assume()
takes more and more time the bigger the assumptions()
data base is. This causes a lot of slowdowns when e.g. declaring variables with a domain
argument. Nils Bruin suggested that this is due to excessive interactions with the Maxima library and Ralf Stephan suggested that the assumptions could be cached and only sent to Maxima when needed, to speed up the process.
Change History (17)
comment:1 Changed 21 months ago by
 Cc rws added
comment:2 Changed 21 months ago by
 Cc egourgoulhon added
comment:3 Changed 21 months ago by
 Description modified (diff)
The "more persistent domains() database" exists already in part as GiNaC/Pynac info flags that are set in parallel to Maxima's assumptions. They can be queried with ex.is_real() etc...What is not saved in Pynac are less elementary assumptions like x>1, y+z==pi. Now instead of caching all assumptions in a database (either Python or C++) and sending to Maxima on demand in bulk, another possibility could be, as you say, to just remove the assume calls on variable creation because they are all elementary assumptions. Then when Maxima needs them for integration, solving etc take the information from Pynac and do assumes for just those variables that are needed. Am I missing something?
comment:4 followup: ↓ 5 Changed 20 months ago by
@rws, that would be awesome. My current workaround is not to pass the domain information to var() at all, but just save it in a separate assumptions database, in case it is needed at some point. Unfortunately, this also prevents the domain information from being passed to GiNaC/Pynac. Removing the assume() calls during variable creation would be a neater way of going about this. The problem that assume() takes longer and longer the more assumptions have already been passed, could then be approached independently. Should I try to remove the assume() calls, run the doctests and try to create a patch, or could someone else? Sorry about my ignorance regarding the development processes for SageMath.
comment:5 in reply to: ↑ 4 ; followup: ↓ 9 Changed 20 months ago by
Replying to schymans:
...Should I try to remove the assume() calls, run the doctests and try to create a patch, or could someone else?
Only removing the calls will make all doctests fail that rely on different variable domains than complex with operations that use Maxima, like integration and solving equations. So additional code is needed. I'll look into it. Of course you can change it in your local copy if you don't need these operation. However it *is possible that other things break.
comment:6 followup: ↓ 7 Changed 20 months ago by
I thought that in most doctests relying on other variable domains in maxima those would be passed as assume() anyway. I haven't really seen the var(..., domain=...)
in use. Is there a way to search all doctests for domain=
?
comment:7 in reply to: ↑ 6 ; followup: ↓ 14 Changed 20 months ago by
Replying to schymans:
I thought that in most doctests relying on other variable domains in maxima those would be passed as assume() anyway. I haven't really seen the
var(..., domain=...)
in use. Is there a way to search all doctests fordomain=
?
ralf@ark:~/sage> grep recursive l 'sage: .*var(.*domain=' src/sage/ grep 'py$\pyx$' src/sage/misc/functional.py src/sage/symbolic/ring.pyx src/sage/symbolic/assumptions.py src/sage/symbolic/expression.pyx src/sage/tensor/modules/free_module_tensor.py src/sage/tensor/modules/free_module_alt_form.py src/sage/tensor/modules/comp.py src/sage/geometry/riemannian_manifolds/parametrized_surface3d.py src/sage/functions/other.py src/sage/functions/hyperbolic.py src/sage/functions/log.py src/sage/functions/trig.py src/sage/calculus/wester.py src/sage/calculus/var.pyx
comment:8 Changed 20 months ago by
OK, the doctests do indeed use domain=
in var()
instead of assume()
on many occasions, so this would not work. It is a bit strange, though, to populate the assumptions()
data base through the var()
call, as assumptions()
can be modified and deleted at any time, whereas the GiNaC domain setting is persistent. I think it would be cleaner if all code passed to maxima was accompanied by its own assumptions, which would be partly derived from the GiNaC variable properties and partly userdefined. Would this require changing every single method that calls maxima?
comment:9 in reply to: ↑ 5 Changed 20 months ago by
comment:10 Changed 20 months ago by
 Branch set to u/rws/cache_assumptions_and_only_send_to_maxima_when_needed
comment:11 followup: ↓ 12 Changed 20 months ago by
 Commit set to 10c31a1971edb7c32fd833d9a497a7ade393471f
 Component changed from symbolics to performance
 Dependencies set to pynac0.7.9
 Type changed from task to enhancement
The first commit prevents calls to Maxima, so it should result in a speedup. There is however a bug in Pynac (fixed in 0.7.9) that prevents it from working correctly. With the fix a few doctests fail, so this needs the planned injection of variable domains (EDITED).
New commits:
10c31a1  23138: don't call Maxima with new symbols

comment:12 in reply to: ↑ 11 ; followup: ↓ 13 Changed 20 months ago by
Following on from https://groups.google.com/forum/#!topic/sagedevel/A8ZzSKvYsA, assume()
sends assumptions both to maxima and Pynac. In addition, assume()
tests if the assumptions are consistent, which is probably not needed when we define a new variable. So, instead of calling assume()
in var()
maybe we should just replace it by the same as is done in assume()
(https://github.com/sagemath/sage/blob/master/src/sage/symbolic/expression.pyx#L1785), i.e. maxima.assume()
.
Therefore, instead of removing send_sage_domain_to_maxima as in the patch, I would propose to substitute maxima.assume()
for all assume()
calls in send_sage_domain_to_maxima()
(https://github.com/sagemath/sage/blob/master/src/sage/symbolic/ring.pyx#1017). Did I miss something?
Replying to rws:
The first commit prevents calls to Maxima, so it should result in a speedup. There is however a bug in Pynac (fixed in 0.7.9) that prevents it from working correctly. With the fix a few doctests fail, so this needs the planned injection of variable domains (EDITED).
New commits:
10c31a1 23138: don't call Maxima with new symbols
comment:13 in reply to: ↑ 12 Changed 20 months ago by
Replying to schymans:
Therefore, instead of removing send_sage_domain_to_maxima as in the patch, I would propose to substitute
maxima.assume()
for allassume()
calls insend_sage_domain_to_maxima()
(https://github.com/sagemath/sage/blob/master/src/sage/symbolic/ring.pyx#1017). Did I miss something?
What takes the time with inconsistency checking is maxima.assume()
so there would be no difference.
comment:14 in reply to: ↑ 7 Changed 20 months ago by
Replying to rws:
Replying to schymans:
I thought that in most doctests relying on other variable domains in maxima those would be passed as assume() anyway. I haven't really seen the
var(..., domain=...)
in use. Is there a way to search all doctests fordomain=
?
Just to mention that var(..., domain='real', ...)
is used in the code (not doctest part) for manifolds (specifically real manifolds); see line 1449 of src/sage/manifolds/chart.py
.
comment:15 Changed 20 months ago by
 Dependencies changed from pynac0.7.9 to #23325
comment:16 Changed 18 months ago by
 Status changed from new to needs_review
Setting to review in order to get a patchbot assessment, now that the abovementioned Pynac fix is in develop. Please set back afterwards.
comment:17 Changed 18 months ago by
 Status changed from needs_review to needs_work
See patchbot log.
Thanks. The Author field is for the developer, you're the Reporter as you can see top left.