CUPTI samples, including event_sampling, callback_event etc. show that all CUPTI APIs used for profiling have to be associated with some context. i.e., profiling is done for a particular application. Is it not possible to use them independently?
Is it not possible to set the counters for particular event for a certain period of time without specifying the context? So that any application that runs during the period, registers the event value in the counter and counter be read at any time instant during this period.
Context handles are required as those operations are configured/controlled independently for each CUDA context.
BTW, CUPTI is designed to be an API for tools venders, not for end CUDA developers (though power users can likely use it to create custom solutions). If you are a CUDA developer you should look at nsight, nvvp, and nvprof.