I redid the benchmark in a non-virtual (VMWare) environment, and the clients did not crash once, even though I tested extensively. I assume therefore that these "AccessViolationException"s are either
- due to the VMWare environment (I have seen somewhat similar issues in other virtual environments before), or
- due to a change I made to the code mid-test trying to increase performance by making the Execution object instance a per-thread instance (rather than sharing the same instance as before, which should be perfectly OK though, if I recall the docs correctly in terms of thread-safety)
I do however believe that it is the first case, because I did quite a number of tests before that code change. So I will mark this as question as answered.