[petsc-users] how to get full trajectory from TS into petsc binary

[Hong Zhang]

Hi Barry,

Actually the TSTrajectory object can save both the solution and the corresponding time information into binary files. Although it is designed specifically for adjoint checkpointing, it does have the side benefit to assist in visualization of simulation trajectories. For adjoint checkpointing, not only solution is saved, but also stage values are saved. So I created a new TSTrajectory type particularly for visualization purpose, which can be access at my branch hongzh/tstrajectory_visualization. 

One can enable this feature using the command line options 

-ts_save_trajectory 1 -tstrajectory_type visualization

The full trajectory will be saved into a folder with multiple files, one file corresponding to one time step. Then we can use a MATLAB script, such as PetscBinaryRead.m, to read these binary files. But the default script coming with PETSc needs to be modified a little bit. Because the trajectory is always saved as a solution vector, followed by a PetscReal variable, I suggest to use the following code in MATLAB:

if  header == 1211214 % Petsc Vec Object
  %% Read state vector
  m = double(read(fd,1,indices));
  x = zeros(m,1);
  v = read(fd,m,precision);
  x(:) = v;
  %% Read time
  t = read(fd,1,precision);
end

Shri (copied in this email) has successfully used this approach to do visualization. Perhaps we can include this feature in the new release if it is useful to some users.

Hong
   

> On Mar 1, 2016, at 10:32 AM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> 
>> 
>> On Mar 1, 2016, at 12:39 AM, Ed Bueler <elbueler at alaska.edu> wrote:
>> 
>> Barry --
>> 
>> Will try it.
>> 
>>> ... since, presumably, other more powerful IO tools exist that would be used for "real" problems?
>> 
>> I know there are tools for snapshotting from PETSc, e.g. VecView to .vtk.  In fact petsc binary seems fairly convenient for that.  On the other hand, I am not sure I've ever done anything "real".  ;-)
>> 
>> Anyone out there:  Are there a good *convenient* tools for saving space/time-series (= movies) from PETSc TS?  I want to add frames and movies from PETSc into slides, etc.  I can think of NetCDF but it seems not-very-convenient, and I am worried not well-supported from PETSc.  Is setting up TS with events (=TSSetEventMonitor()) and writing separate snapshot files the preferred scalable usage, despite the extra effort compared to "-ts_monitor_solution binary:foo.dat"?
> 
>   Ed,
> 
>    As I said in my previous email since we don't have a way of indicating plain double precision numbers in our binary files it is not possible to put both the vectors and the time steps in the same file without augmenting our file format.
> 
>  Barry
> 
> 
> 
>> 
>> Ed
>> 
>> 
>> On Mon, Feb 29, 2016 at 8:53 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>> 
>>  Ed,
>> 
>>   I have added a branch barry/feature-ts-monitor-binary  that supports -ts_monitor binary:timesteps that will store in simple binary format each of the time steps associated with each solution. This in conjugation with -ts_monitor_solution binary:solutions will give you two files you can read in. But note that timesteps is a simple binary file of double precision numbers you should read in directly in python, you cannot use PetscBinaryIO.py which is what you will use to read in the solutions file.
>> 
>>  Barry
>> 
>> Currently PETSc has a binary file format where we can save Vec, Mat, IS, each is marked with a type id for PetscBinaryIO.py to detect, we do not have type ids for simple double precision numbers or arrays of numbers. This is why I have no way of saving the time steps in a way that PetscBinaryIO.py could read them in currently. I don't know how far we want to go in "spiffing up" the PETSc binary format to do more elaborate things since, presumably, other more power IO tools exist that would be used for "real" problems?
>> 
>> 
>>> On Feb 29, 2016, at 3:24 PM, Ed Bueler <elbueler at alaska.edu> wrote:
>>> 
>>> Dear PETSc --
>>> 
>>> I have a short C ode code that uses TS to solve  y' = g(t,y)  where y(t) is a 2-dim'l vector.  My code defaults to -ts_type rk so it does adaptive time-stepping; thus using -ts_monitor shows times at stdout:
>>> 
>>> $ ./ode -ts_monitor
>>> solving from t0 = 0.000 with initial time step dt = 0.10000 ...
>>> 0 TS dt 0.1 time 0.
>>> 1 TS dt 0.170141 time 0.1
>>> 2 TS dt 0.169917 time 0.270141
>>> 3 TS dt 0.171145 time 0.440058
>>> 4 TS dt 0.173931 time 0.611203
>>> 5 TS dt 0.178719 time 0.785134
>>> 6 TS dt 0.0361473 time 0.963853
>>> 7 TS dt 0.188252 time 1.
>>> error at tf = 1.000 :  |y-y_exact|_inf = 0.000144484
>>> 
>>> I want to output the trajectory in PETSc binary and plot it in python using bin/PetscBinaryIO.py.  Clearly I need the times shown above to do that.
>>> 
>>> Note "-ts_monitor_solution binary:XX" gives me a binary file with only y values in it, but not the corresponding times.
>>> 
>>> My question is, how to get those times in either the same binary file (preferred) or separate binary files?  I have tried
>>> 
>>> $ ./ode -ts_monitor binary:foo.dat    # invalid
>>> $ ./ode -ts_monitor_solution binary:bar.dat   # no t in file
>>> $ ./ode -ts_monitor_solution binary:baz.dat -ts_save_trajectory   # no t in file
>>> 
>>> without success.  (I am not sure what the boolean option -ts_save_trajectory does, by the way.)
>>> 
>>> Thanks!
>>> 
>>> Ed
>>> 
>>> PS Sorry if this is a "RTFM" question, but so far I can't find the documentation.
>>> 
>>> 
>>> --
>>> Ed Bueler
>>> Dept of Math and Stat and Geophysical Institute
>>> University of Alaska Fairbanks
>>> Fairbanks, AK 99775-6660
>>> 301C Chapman and 410D Elvey
>>> 907 474-7693 and 907 474-7199  (fax 907 474-5394)
>> 
>> 
>> 
>> 
>> -- 
>> Ed Bueler
>> Dept of Math and Stat and Geophysical Institute
>> University of Alaska Fairbanks
>> Fairbanks, AK 99775-6660
>> 301C Chapman and 410D Elvey
>> 907 474-7693 and 907 474-7199  (fax 907 474-5394)

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20160301/1a78e8ea/attachment.html>