LB inertialess tracers bugs

[Jean-Noël Grad]

Hi,

Severe bugs were discovered in the CPU and GPU implementations of 
inertialess tracers virtual sites.

1. CPU implementation: The integrator would not update the z-component 
of the tracer position. This affects the python and 4.2 branches of 
ESPResSo since February 2022. The python branch has been patched today.

2. GPU implementation: Tracers on non-root MPI ranks would be silently 
ignored by the CUDA kernels, and would have a constant velocity, either 
0 if the particle never visited the fluid domain of MPI rank 0, or the 
last known velocity if the particle was once on the fluid domain of MPI 
rank 0. This bug affected all ESPResSo versions. GPU simulations 
executed with only 1 MPI rank or without MPI were not affected. The 
python branch now raises an exception when attempting to use LB GPU with 
2 or more MPI ranks.

Due to the nature of these bugs, it is difficult to determine if a 
simulation was affected by looking only at the trajectory file or the 
output of an autocorrelation function. We recommend that you look at the 
metadata of your simulation results, specifically the number of MPI 
ranks for a LB GPU simulation or the commit revision for a LB CPU 
simulation, to determine whether the results can be trusted.

You can find a patch for 4.2.1 on GitHub [1]. It has been submitted to 
Fedora 38 [2], Fedora 39 [3] and openSUSE [4].

Best regards,
JN

Links:
[1] 
https://patch-diff.githubusercontent.com/raw/espressomd/espresso/pull/4714.diff
[2] 
https://src.fedoraproject.org/rpms/espresso/c/4f717b07ba7662537ccdc9730d45fd67b2110a54?branch=f38
[3] 
https://src.fedoraproject.org/rpms/espresso/c/93d3c0cde43344b8d0ae2609b063c1b015c40a4e?branch=f39
[4] https://build.opensuse.org/request/show/1083076