When it multiplexes, how accurate are the outputs? Internally, is each counter multiplied by a factor to compute what the actual output is? What happens when the 2 counters must be simultaneously incremented?
I remember prof said they weren't accurate if >4 and should not be used. I can't recall the reasoning.
llcoolj
multiplexing runs all the events for only a little bit, and then provides an estimate based off the time an event is running, scaled up to the total time enabled, which can cause for blind spots that may make the results very inaccurate
ant
The number of different performance counters available for different processor architectures is listed here: https://en.wikipedia.org/wiki/Hardware_performance_counter.
As we saw later in the lecture, out-of-order execution also poses a problem to the accuracy of measurements that we can obtain using these counters.
When it multiplexes, how accurate are the outputs? Internally, is each counter multiplied by a factor to compute what the actual output is? What happens when the 2 counters must be simultaneously incremented?
I remember prof said they weren't accurate if >4 and should not be used. I can't recall the reasoning.
multiplexing runs all the events for only a little bit, and then provides an estimate based off the time an event is running, scaled up to the total time enabled, which can cause for blind spots that may make the results very inaccurate
The number of different performance counters available for different processor architectures is listed here: https://en.wikipedia.org/wiki/Hardware_performance_counter.
As we saw later in the lecture, out-of-order execution also poses a problem to the accuracy of measurements that we can obtain using these counters.