Solved: Analysis Result in Nanoseconds with VTune Amplifier XE

CWhit10 · ‎09-09-2016

Hi,

Is there any way to get the analysis result (i.e. CPU time, overhead, spin time, ...) of Basic Hotspots in terms of nanoseconds?

Thank you.

Alexandra_S_Intel · ‎09-21-2016

Hi, Claudia,

No, I don't believe there's any way to have your metrics display in nanoseconds directly. However, as the metrics are given in seconds, you can approximately convert the numbers yourself by multiplying by 1,000,000,000 (e.g. 5.103s = 5,103,000,000 nanoseconds). That being said, I don't think that will help you; you're probably looking for more accurate numbers.

There are two very good reasons why VTune Amplifier would not display metrics in nanoseconds. The first being, as shown in my example, it would bloat the numbers considerably. The second being that I am almost completely certain that VTune isn't actually accurate to the nanosecond.

While I personally can't explain the details behind the inner workings of a basic hotspots analysis, I do know that the hardware event based analysis types like advanced hotspots use sampling, and therefore would not be accurate down to the nanosecond. This page has more information on sampling and how it works if you're interested. The short version is that VTune periodically interrupts the program every few instructions and "writes down" what happened on the currently tracked events since the last interruption. Otherwise, data collection would be incredibly slow and everything would grind to a halt; it wouldn't be feasible. It also periodically switches out the currently tracked events, only able to count four at a time due to hardware limitations. On a large scale, this is fine, but it's not accurate down to such short intervals.

View solution in original post

Alexandra_S_Intel · ‎09-21-2016

Hi, Claudia,

No, I don't believe there's any way to have your metrics display in nanoseconds directly. However, as the metrics are given in seconds, you can approximately convert the numbers yourself by multiplying by 1,000,000,000 (e.g. 5.103s = 5,103,000,000 nanoseconds). That being said, I don't think that will help you; you're probably looking for more accurate numbers.

There are two very good reasons why VTune Amplifier would not display metrics in nanoseconds. The first being, as shown in my example, it would bloat the numbers considerably. The second being that I am almost completely certain that VTune isn't actually accurate to the nanosecond.

While I personally can't explain the details behind the inner workings of a basic hotspots analysis, I do know that the hardware event based analysis types like advanced hotspots use sampling, and therefore would not be accurate down to the nanosecond. This page has more information on sampling and how it works if you're interested. The short version is that VTune periodically interrupts the program every few instructions and "writes down" what happened on the currently tracked events since the last interruption. Otherwise, data collection would be incredibly slow and everything would grind to a halt; it wouldn't be feasible. It also periodically switches out the currently tracked events, only able to count four at a time due to hardware limitations. On a large scale, this is fine, but it's not accurate down to such short intervals.

Dmitry_P_Intel1 · ‎09-22-2016

Hello Claudia,

Basic Hotspots analysis sampling interval minimum is limited by system timer resolution ability that should be 10ms on Windows and 1ms on Linux.

Alex correctly pointed that you need Advanced Hotspots to set up finer grain sampling interval.

BTW - what is the reason why you want to have nanosecond resolution?

Thanks & Regards, Dmitry