GitLab

This adds support for configuring and running simulations using the ns3
end-to-end framework.

Add all TCP congestion control algorithms that are currently available
in the kernel source tree as loadable modules.

This temporarily drops support for DDIO and the L3 cache.

The doc string says to return 1 in the case of an unnecessary sync, which the implementation actually doesn't do. It instead considers an unnecessary sync as success, which is generally also the intended behavior.

We already have a similar and much more detailed section in howto.rst. Further, the included run.py usage print will often be out-of-date and manually updating it every time isn't maintainable.

To manage external links such as for Python Modules, we are now using the extension sphinx.ext.extlinks to define all of them in the file doc/external_links.py.

ExpEnv hd_path and hd_path_raw use absolute path unmodified if given as param

Before this, if registry or tag got set, we would still build all images
starting from the docker hub regsitry version and latest tag.

This makes it easier for images that build on ours to use tools like packer
without having to specify the full path

Update to newer ns-3 version and adapt code for changes (mostly
renamed files and command line parameters) in the ns-3 submodule.

besides main time, print in and out timestamp for the ports' underlying SimBricks interfaces

besides main time, we now also print in and out timestamp of the underlying SimBricks interfaces

fix container, Python version, path to sphinx' conf.py and add all
required packages

Switch cloud-init data source from nocloud-net to nocloud, since
noncloud-net is now marked as deprecated and same functionality is
provided by nocloud.

Switch to an older release of the ubuntu minimal cloud image, because
the most recent version (20240222) introduced a bug that prevents the
base image from building.

Check the length of socket paths in SimbricksBaseIfListen and
SimbricksBaseIfConnect. If the length exceeds the size of sun_path in
the sockaddr_un struct, we fail with a meaningful error message.

This resolves #42.

This change is mostly about the default semantics of asyncio.wait(). It
doesn't forward any exceptions that are raised by any of the coroutines
it waits for. This effectively causes exceptions to be swallowed. In
contrast, asyncio.gather() forwards exceptions to its calling coroutine.

The reason we didn't spot this earlier is that exceptions were swallowed
by the call to `asyncio.wait()` in
`runners.py:ExperimentBaseRunner.terminate_collect_sims()` when invoking
`int_term_kill()`. Instead, we now use `asyncio.gather()`, which
forwards exceptions to the calling coroutine.

Currently --repo, which is used to find simulators etc., defaults to '..',
which only works when executing in the 'experiments' directory in the simbricks
repo. Instead we make this relative to the run.py-file's path.

So far, simulator instances not connected to any other instances were
skipped accidentally when creating the graph.

Currently, when a checkpoint is created in gem5 using the KVM CPU, the
simulated tick when the checkpoint is taken will end up in the few
seconds or more when the KVM CPU is slow. When the checkpoint is
afterwards restored, gem5 will continue the simulation at this tick.
This means that other, synchronized simulators first have to catch up,
essentially just synchronizing without performing any useful work.
Depending on their speed, this can take a lot of time and is especially
costly with a RTL simulator in the loop. To avoid this we now, by
default, edit the to be restored gem5 checkpoint to start at tick 0
instead.

Add a cleanup script for ns-3 similar to those for qemu and gem5.

This solves #9.

When we switched to a newer kernel version in c64d3d0e, the
cleanup_images.sh script no longer removed the linux kernel source
folder, since it still used the old version number. Now, I use a regex
to avoid this issue in the future.