Unfortunately, the Comsol API does not support running more than one client at a time, i.e. within the same Java program. Meanwhile, JPype cannot manage more than one Java virtual machine within the same Python process. If it could, it would be easy to work around Comsol’s limitation. (There is an alternative Java bridge, pyJNIus, which is not limited to one virtual machine, but then fails in another regard: A number of Java methods exposed by Comsol are inexplicably missing from the Python encapsulation.)
Therefore, if several simulations are to be run in parallel, distributed over independent processor cores in an effort to achieve maximum speed-up of a parameter sweep, they have to be started as separate Python subprocesses. Refer to section “Multiple processes” for a demonstration.
Additionally, there are some known, but unresolved issues with JPype’s
shutdown of the Java virtual machine. Notably, pressing Ctrl+C
to interrupt an ongoing operation will usually crash the Python session.
So do not rely on catching
exceptions in application code.
The Comsol API offers two distinct ways to run a simulation session on
the local machine. We can either start a “stand-alone” client, which
does not require a Comsol server. Or we first start a server and then
have a “thin” client connect to it via a loop-back network socket. The
first approach is more lightweight and more reliable, especially on
Windows, as it keeps everything inside the same process. The second
approach is slower to start up and relies on the inter-process
communication to be robust, but would also work across the network,
i.e., for remote sessions where the client runs locally and delegates
the heavy lifting to a server running on another machine. If we
Client class without providing a value
for the network port, it will create a stand-alone client. Otherwise
it will run as a thin client in client–server mode.
On Linux and macOS, however, the stand-alone mode does not work out of
the box. This is due to a limitation of Unix-like operating systems
and explained in more detail in GitHub issue #8. On these platforms,
if all you did was install MPh, starting the client in stand-alone mode
will raise a
java.lang.UnsatisfiedLinkError because required external
libraries cannot be found. You would have to add the full paths of
certain shared-library folders to an environment variable named
LD_LIBRARY_PATH on Linux and
DYLD_LIBRARY_PATH on macOS.
For example, for an installation of Comsol 6.0 on Ubuntu Linux, you
would add the following lines at the end of the shell configuration
# Help MPh find Comsol's shared libraries. export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:\ /usr/local/comsol60/multiphysics/lib/glnxa64:\ /usr/local/comsol60/multiphysics/ext/graphicsmagick/glnxa64:\ /usr/local/comsol60/multiphysics/ext/cadimport/glnxa64
On macOS, the root folder is
The folder names depend on the installed Comsol version and will have
to be adapted accordingly.
Requiring the environment variable to be set correctly limits the possibility of selecting a specific Comsol version from within MPh, as adding multiple installations to that search path will lead to name collisions. One could work around the issue by wrapping a Python program using MPh in a shell script that sets the environment variable before starting the Python process. That’s effectively what Comsol itself does to launch its GUI on one of these platforms. Or we could have a Python program that sets the environment variable and then runs MPh in a second Python subprocess. Clearly, none of this is ideal. Starting the client should work without any of these detours.
mph.start() exists to navigate these platform
differences. On Windows, it starts a stand-alone client in order to
profit from the better start-up performance. On Linux and macOS, it
creates a local session in client–server mode so that no shell
configuration is required up front. This behavior is reflected in the
'session', accessible via
mph.option(), which is set to
'platform-dependent' by default.
It could also be set to
start() in order to override the default behavior.
Performance in client–server mode is noticeably worse in certain
scenarios, not just at start-up. If functions access the Java API
frequently, such as when navigating the model tree, perhaps even
mph.tree() does, then client–server mode can be
slower by a large factor compared to a stand-alone client. Rest assured
however that simulation run-times are not affected.
Conversely, setting up stand-alone mode on Linux or macOS is also not a robust solution. Image exports, for example, are known to crash due to some conflict with external libraries. As opposed to Windows, where this works reliably.