如何知道进程正在使用哪个处理器？

Question

In a multithreaded environment, I want to exactly know which processor is being used by my process.

I looked into the source code of top and htop. The problem with the top/htop is that it gets the total time (from /proc/stat and /proc/[pid]/stat ) and divides it by the number of CPUs irrespective of the load that is distributed among the other processors.

If the total %CPU is 350% theoretically we know at least 4 processors are being used. The top is generalizing this and doesn't "accurately" quantify the per processor usage.

350% can imply in different ways:

Processor 1: 100%
Processor 2: 75%
Processor 3: 75%
Processor 4: 100%

Processor 1: 75%
Processor 2: 100%
Processor 3: 100%
Processor 4: 75%

The /proc/[pid]/stat gives the last processor used but doesn't which processors are being used by a process.

Any idea how to accurately capture which processor is being by the process?

My end goal is to find out how efficiently my algorithm is using the machine.

Answer 1

As said in the comments of your question, there's no point in asking the system where your process is executed as the scheduler can change it before you even get the reply.

In a Posix environment you can use pthread_setaffinity_np() to force your code to only execute on certain threads/cores although I'm not sure that will be too helpful to you find out the efficiency of your algorithm.

When I want to profile a multi-threaded algorithm like that I'd probably use something like gnome-system-monitor or the KDE system monitor (gotop would work too) to print out CPU usage graphs so I can compare the system load before and after starting my program, it can at least help find out if there's 1 thread completely left alone, not very granular but for the first pass it's enough.

Once I'm too close to 100%/the program is too fast to see whether I'm using all available resources or not, I'd usually do something like

auto timeBefore = std::chrono::high_resolution_clock::now();
my_algo();
auto timeAfter = std::chrono::high_resolution_clock::now();
std::chrono::duration<double> completeTime = timeAfter-timeBefore;
std::cout << "Took: " << completeTime.count() << "s" << std::endl;

so I can do some benchmarking cause after all, execution time is usually more important than having all cores pinned to 100%