long-running daemons and cgroups

Vendor daemons are preventing us from implementing cgroups for process accounting.

We would like to implement cgroups for process accounting in our LSF
cluster to resolve a long standing issue of processes escaping from
LSF control.

We have several vendor daemons that start as part of a user job that
will either run forever or until some internal inactivity timeout. In some
cases, the daemons are shared among users. The first job will start
the daemon and subsequent jobs will use that process rather than starting
its own. The daemons run on a well known port, so only one can run at a time.

With cgroup accounting enabled, the job that started the daemon will
remain in a RUN state until the job is killed or the daemon reaches
its inactivity timeout. In some cases, it might not reach the timeout
because of other jobs keeping the process active.

One possible solution we are looking into is starting the shared daemons
at boot time from a non-LSF user account. That might solve one issue,
but we still have daemons that start for each user and don't appear to
timeout. I am talking to the vendor about the behavior to see if there is
anything we can do to ensure those daemons only last as long as the job.

If we are unable to resolve the issue with vendor changes or vendor
guidance, I've considered options such as a post-exec that would kill
the processes belonging to the job's cgroup. We can't blindly kill
everything in the cgroup because that would kill res or other LSF daemons
resulting in the jobs always exiting with an EXIT status rather than DONE.

Has anyone else faced this issue? Are there alternate solutions to the
problem or ideas on how to solve it? Could LSF be modified to exclude
certain processes from the cgroup?