#2101: Add Lean_public and Proca thorns to the ET --------------------------------------------------------+------------------- Reporter: miguel.zilhao.nogueira@… | Owner: sbrandt Type: enhancement | Status: assigned Priority: major | Milestone: ET_2018_08 Component: EinsteinToolkit thorn | Version: development version Resolution: | Keywords: Lean_public --------------------------------------------------------+------------------- Changes (by sbrandt):

* owner: => sbrandt * status: confirmed => assigned

#2101: Add Lean_public and Proca thorns to the ET ---------------------------------------+--------------------------------- Reporter: miguel.zilhao.nogueira@… | Owner: Steven R. Brandt Type: enhancement | Status: assigned Priority: major | Milestone: ET_2018_08 Component: EinsteinToolkit thorn | Version: development version Resolution: | Keywords: Lean_public ---------------------------------------+---------------------------------

Comment (by Peter Diener):

Here is my review of the "Proca" suite of thorns that are proposed for inclusion in the Einstein Toolkit. Proca consists of ProcaBase, Proca_simpleID, TwoPunctures_KerrProca and ProcaEvolve and NPScalars_Proca. The code is nicely written, but mainly lack some tests and documentation.

Questions, comments and recommendations regarding Proca overall:

There are two tests in ProcaEvolve of which one uses Lean and the other ML_BSSN while otherwise appear to be the same. Whereas the Ai and Ei variables agree quite well, there seems to be significant differences between Aphi and Zeta. Is this expected? If so, why?

In ProcaEvolve there is another parameter file in the test directory (ML_BSSN_Aphi_mu0.4_c100.1_r06_S0.0_4th.par) that does not have any associated output. If this was meant to be a test, output should be provided. Otherwise it should be moved to the example parameter file directory.

NPScalars_Proca, Proca_simpleID and TwoPunctures_KerrProca have no tests of their own. Proca_simpleID is tested by the tests in ProcaEvolve, but NPScalars_Proca and TwoPunctures_KerrProca should definitely be covered by tests. You may also consider making a standalone test for Proca_simpleID.

None of the thorns in the Proca arrangement have any documentation of their own. As there is a paper describing the formulation, it is okay to refer to the paper, but there should be some description of the variables and parameters used in the code and how they relate to the physics and equations.

It would be nice if all he parameter files for the runs in the paper could be made available as examples of how to use the thorn. Currently there are only two example parameter files available and it's not immediately clear if they have been used for the paper. There could be a comment at the top of the parameterfiles indicating what section and/or figure in the paper they were used for.

Questions, comments and recommendations specifically to ProcaBase:

In parameter.ccl the comment to the value of the parameter mu says: "any non-negative number". However, the range allows for zero and in fact the default value is zero. Either the comment should be changed or the range and default value should be changed.

Are the integer power for the fall off rate for the outgoing boundary conditions for the Proca variables not known? Or can they be different for different physical systems? If they are known constants, should they be parameters at all? Also could the power be different for different components of the E and A vectors. If not, there's no need to make these vector parameters.

I propose to give the file (src/Basegrid.F90) with the symmetry registering routine a more descriptive name.

It would be nice to have a proper (but short) thorn documentation to relate the variables with the paper.

Questions, comments and recommendations specifically to Proca_simpleID:

The Simple Initial Data solution in the paper is a spherical symmetric solution. The parameters for Proca_simpleID on the other hand seem to allow for a superposition of 2 such solutions separated by a non-zero distance. As the parameter for the separation can not be zero (imposed by the parameter range) the initial data thorn will thus never produce initial data that is spherically symmetric around the origin of the cactus grid. Would it make sense for the range of par_b to include zero in order to allow for that. This is something worth explaining in the (missing) documentation.

Questions, comments and recommendations specifically to TwoPunctures_KerrProca:

It looks like the code implements initial data with a Y_11 angular dependence in addition to Y_00 and Y_10 mentioned in the paper. This should be in the documentation.

There's a comment in Equations.c that some of the equations were copied from Mathematica. Could that mathematica notebook be made available in the thorn as well?

In the future it might be worth considering some way of merging this back into the original TwoPunctures, so we don't have multiple thorns implementing the same spectral solver infrastructure.

There should be tests for all the different types of angular profiles for the initial data.

The discussion in the README should be converted into a proper thorn documentation.

Questions, comments and recommendations specifically to ProcaEvolve:

Need a thorn documentation to relate variables to the paper and to explain the use of parameters.

I don't know how much performance is a concern. If only a tiny amount of time is spent in Proca compared to BSSN (or other things) this is obviously not important. However, if the issues with the code, that prevents the compiler from optimizations, makes the code slow enough that the time spent in Proca is comparable to the time spent in BSSN it might be worth it to fundamentally rewrite the code. One of the problems, is the many nested loops over tensor indices. These prevents the compiler from vectorizing most of the RHS code (I checked this by looking at compiler vectorization output from the Intel compiler). As the loop lengths are just 3, the vectorizer deems that vectorization will be ineffecient and hence doesn't do it. On modern CPUs with vector lengths of 4 or 8 (in double precision), efficiently using the vector units is really important for good performance. However, this is not an issue (if there really is a problem with performance) that should prevent inclusion in the ET, but just something to think about for later.

#2101: Add Lean_public and Proca thorns to the ET ---------------------------------------+--------------------------------- Reporter: miguel.zilhao.nogueira@… | Owner: Steven R. Brandt Type: enhancement | Status: assigned Priority: major | Milestone: ET_2018_08 Component: EinsteinToolkit thorn | Version: development version Resolution: | Keywords: Lean_public ---------------------------------------+---------------------------------

Comment (by miguel.zilhao.nogueira):

Replying to [comment:4 Peter Diener]:

many thanks for your input and comments! below we try to address your points.

Questions, comments and recommendations regarding Proca overall:

There are two tests in ProcaEvolve of which one uses Lean and the other

ML_BSSN

while otherwise appear to be the same. Whereas the Ai and Ei variables

agree

quite well, there seems to be significant differences between Aphi and

Zeta.

Is this expected? If so, why?

on my machine Aphi gives very similar values for the evolutions with Lean and ML_BSSN; Zeta does appear a bit different, but this quantity is the constraint damping variable and has no physical meaning, so this may just be due to the different ways the gauge is handled in Lean in ML_BSSN...

In ProcaEvolve there is another parameter file in the test directory (ML_BSSN_Aphi_mu0.4_c100.1_r06_S0.0_4th.par) that does not have any

associated

output. If this was meant to be a test, output should be provided.

Otherwise

it should be moved to the example parameter file directory.

it has been moved to the par folder, thanks.

NPScalars_Proca, Proca_simpleID and TwoPunctures_KerrProca have no tests of their own. Proca_simpleID is tested by the tests in ProcaEvolve, but NPScalars_Proca and TwoPunctures_KerrProca should definitely be covered

by

tests. You may also consider making a standalone test for

Proca_simpleID.

added respective tests.

None of the thorns in the Proca arrangement have any documentation of

their

own. As there is a paper describing the formulation, it is okay to refer

to the

paper, but there should be some description of the variables and

parameters

used in the code and how they relate to the physics and equations.

we've added to the ProcaBase interface.ccl file some more description, matching the variables with the notation used in the first paper. once we have the new paper out this can also be updated to reflect the notation therein, if needed. we've also added a simple documentation.tex to ProcaEvolve and TwoPunctures_KerrProca

It would be nice if all he parameter files for the runs in the paper

could be

made available as examples of how to use the thorn. Currently there are

only

two example parameter files available and it's not immediately clear if

they

have been used for the paper. There could be a comment at the top of the parameterfiles indicating what section and/or figure in the paper they

were

used for.

we've added this comment in one of the parfiles. the other two were used for the runs in the upcoming paper; we can add the information as well once the paper is published.

Questions, comments and recommendations specifically to ProcaBase:

In parameter.ccl the comment to the value of the parameter mu says: "any non-negative number". However, the range allows for zero and in

fact the

default value is zero. Either the comment should be changed or the range

and

default value should be changed.

zero is indeed allowed, this is what we meant with "non-negative". does it not imply that zero is permissible?

Are the integer power for the fall off rate for the outgoing boundary conditions for the Proca variables not known? Or can they be different for different physical systems? If they are known constants, should they be parameters at all? Also could the power be different for different components of the E and A vectors. If not, there's no need to make these vector parameters.

i think there have not been any rigorous studies of the boundary conditions of these systems... the default values we are setting are what we expect makes more sense from a naive analysis, but in principle a user may wish to experiment with other values, so we decided to allow for that option.

I propose to give the file (src/Basegrid.F90) with the symmetry

registering

routine a more descriptive name.

it has been renamed to Proca_Symmetries.F90

It would be nice to have a proper (but short) thorn documentation to

relate

the variables with the paper.

would the description introduced to the interface.ccl under ProcaBase suffice? if not, we can certainly add more information if something is not clear...

Questions, comments and recommendations specifically to Proca_simpleID:

The Simple Initial Data solution in the paper is a spherical symmetric solution. The parameters for Proca_simpleID on the other hand seem to allow for a superposition of 2 such solutions separated by a non-zero distance. As the parameter for the separation can not be zero (imposed

by

the parameter range) the initial data thorn will thus never produce

initial

data that is spherically symmetric around the origin of the cactus grid. Would it make sense for the range of par_b to include zero in order to allow for that. This is something worth explaining in the (missing) documentation.

this initial data thorn actually allows for more configurations than the Simple Initial Data one of the paper. by allowing a superposition of 2 solutions, we can also construct the initial data for the head-on collision of two charged black holes with equal charge-to-mass ratios (as in https://arxiv.org/abs/1205.1063).

however, due to how the initial data is constructed, the parameter par_b cannot be exactly zero, and we therefore have excluded that option in the parameter range. choosing a very small value for it (like the default choice) typically works very well at producing spherically symmetric data.

we have added this information to a README file in this thorn.

Questions, comments and recommendations specifically to

TwoPunctures_KerrProca:

It looks like the code implements initial data with a Y_11 angular

dependence

in addition to Y_00 and Y_10 mentioned in the paper. This should be in

the

documentation.

this is correct. the initial data with the Y_11 mode is something that will be featured in the upcoming paper. we've added this note in the documentation.

There's a comment in Equations.c that some of the equations were copied

from

Mathematica. Could that mathematica notebook be made available in the

thorn

as well?

we have no problem with it being made available, we just need to clean up the notebook a bit.

In the future it might be worth considering some way of merging this

back

into the original TwoPunctures, so we don't have multiple thorns

implementing

the same spectral solver infrastructure.

we considered this, but since the equations to be solved changed considerably, we thought it'd be cleaner to implement this in a new thorn. we're certainly open to the possibility of merging this with the original TwoPunctures if deemed possible and useful, though.

There should be tests for all the different types of angular profiles

for the

initial data.

the issue here is that we find that these initial data need a significant number of spectral collocation points, otherwise the algorithm just doesn't converge. this makes some configurations very time-consuming, needing order ~10-15minutes at least to run. this makes them unfeasible for quick testing... we've added one configuration which runs moderately fast (a couple of minutes); we hope this is ok.

The discussion in the README should be converted into a proper thorn documentation.

we have updated the README and added a documentation.tex file with the notes.

Questions, comments and recommendations specifically to ProcaEvolve:

Need a thorn documentation to relate variables to the paper and to

explain

the use of parameters.

we've added a basic documentation.tex mapping the rhs evolution variables to their definition in the paper.

I don't know how much performance is a concern. If only a tiny amount of

time

is spent in Proca compared to BSSN (or other things) this is obviously

not

important. However, if the issues with the code, that prevents the

compiler

from optimizations, makes the code slow enough that the time spent in

Proca is

comparable to the time spent in BSSN it might be worth it to

fundamentally

rewrite the code. One of the problems, is the many nested loops over

tensor

indices. These prevents the compiler from vectorizing most of the RHS

code

(I checked this by looking at compiler vectorization output from the

Intel

compiler). As the loop lengths are just 3, the vectorizer deems that vectorization will be ineffecient and hence doesn't do it. On modern

CPUs with

vector lengths of 4 or 8 (in double precision), efficiently using the

vector

units is really important for good performance. However, this is not an

issue

(if there really is a problem with performance) that should prevent

inclusion

in the ET, but just something to think about for later.

we did not run very thorough checks, but from our limited analysis it seemed that the BSSN metric evolution code still dominates the runtime, so we didn't worry too much about optimizing the Proca evolution. also, one of the motivations for making this code available was for pedagogical purposes, so we thought that having a readable and easy to understand code was preferable. in any case, we are certainly open to write it in a more efficient manner if there are things being done in an obviously non-optimal way.