Compare VODE to VODE90. Determine if a switch to VODE90 is in order. about microphysics HOT 6 CLOSED

The following test is on merger on development branch of Microphysics using unit_test/test_react with the following compilation flags:

make -j NDEBUG=t MPI= OMP= COMP=gfortran NETWORK_DIR=aprox13 EOS_DIR=helmholtz

with INTEGRATOR_DIR=VODE90 or INTEGRATOR_DIR=VODE.

Due to the recent updates to VODE enforcing limits on mass fractions and T, I'm using commit d98ac7843db80894f78a955f62ca6c3ea282209b from September 12, 2017. The AMReX git commit is 266f31c00dfacb03b58a551006dce04ee6cc60a7 from September 18, 2017.

VODE

[dwillcox@merger test_react]((d98ac78...))$ ./main.Linux.gfortran.exe inputs_aprox13
 MPI initialized with            1  MPI processes
 MPI initialized with            1  threads
 
 Initializing Helmholtz EOS and using Coulomb corrections.
 
 small_temp =    100000.00000000000     
 small_dens =    1.0000000000000000     
 Execution time:    156.29828095436096     
 RHS stats:
   min:            6
   avg:         1522
   max:        23965

VODE90

[dwillcox@merger test_react]((d98ac78...))$ ./main.Linux.gfortran.exe inputs_aprox13
 MPI initialized with            1  MPI processes
 MPI initialized with            1  threads
 
 Initializing Helmholtz EOS and using Coulomb corrections.
 
 small_temp =    100000.00000000000     
 small_dens =    1.0000000000000000     
 Execution time:    160.93007707595825     
 RHS stats:
   min:            6
   avg:         1522
   max:        23965

fcompare

fcompare reports no differences in the plotfiles written by test_react with "PLOTFILES AGREE"

All this suggests that on the CPU the two are pretty comparable with a timing difference relative to VODE of 2.9%.

from microphysics.

I've repeated this test 9 more times for a total of 10, and the average times plus or minus a standard deviation for each integrator are as follows:

VODE: 156.2 +/- 1.2 seconds

VODE90: 159.7 +/- 1.3 seconds

Relative difference wrt VODE: 2.25%

Timing spreadsheet: VODE-vs-VODE90-aprox13.pdf

This suggests there is a performance difference on the CPU but it is relatively small.

from microphysics.

Can we get this in a branch and then run the test suite on it to check differences?

from microphysics.

Not quite yet, a few days ago I realized the timing above is for the original VODE90 revisions (on development), before I made the changes needed to make it operable with CUDA Fortran.

When I've run timing between the versions of VODE90 and VODE on the cudadevice branch in serial mode on the CPU I'm getting larger timing differences of ~7% and need to determine why. I should be able to look into this more today.

from microphysics.

On development I see 3% performance slowdown for VODE90 on aprox13, but not identical diffs, rather O(1e-10).

from microphysics.

Closing as we have already done the merge from VODE90 into VODE.

from microphysics.