background
previously, tried to deploy lgsvl in docker swarm, which is failde due to the conflict of host network to run lgsvl and the routing mesh of swarm, as I thought.
http listen on *
why used –network=host, is actually not a have-to, the alternative option is to use "*" as Configure.webHost, instead of localhost nor a special IP address, which lead o HttpListener error:
|
|
then, we can docker run lgsvl without host network limitations.
but still, if run by docker service create, it reports failure: Error initiliazing Gtk+.
Gtk/UI in Unity
when starting lgsvl, it pops the resolution window, which is a plugin of Unity Editor, and implemented with gtk, as explained in last section, which leads to the failure to run lgsvl as service in docker swarm.
the simple solution is to disable resolution selection in Unity Editor.
|
|
then the popup window is bypassed.
ignore publish port
I tried to ignore network host and run directly with routing mesh, but it still doesn’t work. then I remember at the previous blog, when run vkcube or glxgears in docker swarm, it actually does use --network host, so it looks the failure of running lgsvl in docker swarm, is not due to network host, but is due to Gtk/gui. as we can bypass the resolution UI, then directly running as following, works as expected:
|
|
add assets into container
another update is to bind assets from host into lgsvl image, which is stored as sqlite data.db, which is a necessary, as we bypassed the authentication, and the cluster has no access to external Internet.
where is next
in recent two month, had digged into docker swarm to run lgsvl. so far, the main pipeline looks work now, and there are still a lot little fix there.
to run AV simulation in cloud, is a necessary way to test and verify L3+ AV algorithms/products. previous ADAS test is more on each individual feature itself, e.g. ACC, AEB .e.t.c, all of which are easy to define a benchmark test case, and engineers can easily define the test scenarios systemetically. But for L3+, the env status space is infinite in theory, there is no benchmark test cases any more, and at best we can do is statiscally test cases, which requires a huge number of test cases, which is where virutal simulation test in cloud make sense.
from tech viewpoint, the next thing is how to drive the L3+ dev by these simulation tools. and another intesting is the data infrastructure setup