SimGrid
3.19.1
Versatile Simulation of Distributed Systems
|
This section describes jMSG, the Java API to Simgrid. This API mimicks MSG, which is a simple yet somehow realistic interface. The full javadoc is available.
Most of the documentation of the MSG API in C applies directly to the Java bindings (any divergence is seen as a bug that we should fix). MSG structures are mapped to Java objects as expected, and the MSG functions are methods in these objects.
This is documented in the Build the Java bindings Section.
In most cases, you can use the SimGrid bindings as if it was a Java library:
For example:
Any SimGrid simulation (java or not) is usually constituted of several kind of actors or processes (classes extending Msg.Process
) that are deployed over the hosts of the virtual platform. So, your code should declare these actors, plus a Main class in charge of deploying your actors on the platform. Please refer to the examples for details.
Actually, these bindings are not only implemented in Java. They do use the C implementation of SimGrid. This should be transparent as this library is directly included in the simgrid.jar
file but things can still go wrong is several ways.
Error: library simgrid not found
This means that the JVM fails to load the native library. You should try to rebuild the simgrid.jar file as explained above. If it does not help, you can try to use an installed version of the library instead of the one included in the jar. For that, specify the path to the native library in the LD_LIBRARY_PATH
variable (or in the DYLD_LIBRARY_PATH
on Mac OSX).
Add these lines to your ~/.bashrc
file or equivalent to make these settings permanent even after a reboot.
Other errors
When using jMSG, your program can crash for 3 main reasons:
THIS SECTION IS SOMEWHAT OBSOLETE because building a JVM providing coroutines is probably not possible anymore nowadays. If you really need performance for your Java simulation, please contact us.
There is two main motivations to use the coroutine variant of SimGrid Java bindings: it's about 5 times faster than the default thread-based context factory, and the amount of runnable processes is then only limited by the amount of RAM that you have. The drawbacks are that it requires a specific and rather experimental JVM to run, and that this context factory itself remains a bit experimental so far.
You need to get a patched JVM from here (many thanks to Lukas Stadler for this work!).
You can either get a prebuilt binary, or recompile your own JVM. Make sure to get a coro-simple version, as we don't need to serialize nor migrate stacks in SimGrid. You should be able to follow the README.txt
that you'll get in the repository, but here is how we did it, just in case. The instructions are given for a debian or Ubuntu box, but I think you should manage to convert it to your system quite easily. Finally, if you're really stuck, you can get the version compiled by Jonathan Rouzaud-Cornabas from his web page. This version is known to work with SimGrid for sure! http://graal.ens-lyon.fr/~jrouzaud/files/corosimple-linux-amd64-20120914.tgz
Configure the mercurial extensions: Edit ~/.hgrc and paste the following lines. Don't forget to change the /path/to/forest.py to point to where you just downloaded the source.
Forest extension is needed to download the openjdk source code and patches while the mq line is needed to apply the patches. The username is needed at the step "preparing the sources", not sure why.
SimGrid Java will automatically switch to the coroutine context factory if your JVM support it, so you will just need to execute your simulation with the correct JVM. The selected context factory gets displayed automatically.
Note that you may have to adjust the "coro.stacksPerThread" configuration option to run large simulations. The default is 100 and you want to increase it to run more processes.
If you reach the point where the creation of new simulated processes fail with the message "Can't create coroutine object", you may need to increase the relevant system limit with the following command.
The full story is that each coroutine requires two memory maps, and that Linux puts a limit on the total amount of memory maps that each process can manage (by default, this limit is often at 65535). Since the JVM needs a few dozen of such maps on its own (three maps per dynamic library – check /proc/the_pid/maps
if you don't believe it), this is enough to create over 30,000 simulated processes. But to go futher, that limit must be modified.
If you want to make this change permanent on your machine, edit your /etc/sysctl.conf
file. Otherwise, you have to redo it by calling sysctl after each reboot.