Measuring the clock rate of an Intel CPU in C++

I created a short program to directly measure the clock rate of an Intel CPU. It runs, sleeps for a duration specified by the user, and then measures the number of clock cycles and the amount of time that elapsed while it was asleep, then divides the elapsed clock cycles by the elapsed time to get the frequency of the CPU.

• Is the below code up to stuff in terms of style and readability?
• Is there any reason the strategy used and described below won't produce accurate results?

Any feedback is appreciated! If you wish to test out the code, you can download it with git clone https://github.com/firetotherain/CPUHertz.git

#include <thread>
#include <chrono>
#include <cstdio>

//Uses assembly command to get the current value of the cycle "counter"
uint64_t get_cycles()
{
unsigned int lo,hi;
__asm__ __volatile__ ("rdtsc" : "=a" (lo), "=d" (hi));
return ((uint64_t)hi << 32) | lo;
}

//Defines unit of time to measure seconds
typedef std::chrono::duration<double, std::ratio<1,1>> seconds_t;

//In main(int argc, char** argv) this global variable is initialized to argv[0]
const char* program_name;

//Initialized at the start of the program; gets the current time
auto start_time = std::chrono::high_resolution_clock::now();

//Returns the time since the start of the program, measured in seconds
//Has accuracy identical to that of std::chrono::high_resolution_clock
double age()
{
return seconds_t(std::chrono::high_resolution_clock::now() - start_time).count();
}

//Prints the program usage
void PrintUsage()
{
printf("Usage:\n");
printf("%s [measurment duration]\n", program_name);
}
int main(int argc, char** argv)
{
using namespace std::chrono_literals;
program_name = argv[0];
int sleeptime = 100;
switch(argc) {
case 1:
sleeptime = 100;
break;
case 2:
try
{
sleeptime = std::stoi(argv[1]);
}
catch(...)
{
printf("Error: argument was not an integer.");
PrintUsage();
return 1;
}
break;
default:
printf("Error: too many arguments.");
PrintUsage();
return 1;
break;
}
uint64_t cycles_start = get_cycles();
double time_start = age();
uint64_t elapsed_cycles = get_cycles() - cycles_start;
double elapsed_time = age() - time_start;
printf("CPU MHz: %.3f\n", elapsed_cycles / elapsed_time / 1000000.0);
}


There are a number of things that you might employ to improve your program.

Accuracy

Is there any reason the strategy used and described below won't produce accurate results?

Yes! There are many reasons, which are each enumerated separately in items below, followed by more general information about code style and structure.

Understand System Management Mode

The code currently does not account for the possibilty of the processor entering System Management Mode (SMM). One simple way of putting the machine into SMM is to have it go into standby. My laptop is configured to go into standby when I close the lid and come back out when I open it again, so when I did that when running this program, I got this result:

CPU MHz: 2365092677585.790

While it would be lovely to actually have a processor that fast, my laptop does not, so this result is incorrect.

Understand sleep_for

What std::this_thread::sleep_for(delay) does is to sleep for at least delay but it may be longer, depending on what else the operating system and the underlying hardware are doing at the moment. This means that there is a variability in the results of the program. An example of one thousand runs of the program is shown plotted below.
While in this case, most of the results were around the correct value of 2494.225 MHz, the results vary. The alternate way to do this would be to use a hardware-based timer instead and run the timing code as part of the kernel with interrupts disabled.

Understand out-of-order execution and cache effects

The variability of this program as it currently stands is such that it doesn't matter, but when attempting to do precise timing using the RDTSC instruction, it's important to understand that the processor does out-of-order execution. This means that instructions preceding or following the RDTSC instruction may actually be part of the duration of the event you're trying to measure unless precautions are taken. What is generally used is a serializing instruction such as CPUID to eliminate such effects. Intel has a useful whitepaper that describes this in detail and how to apply it.

What follows is more general advice about the coding style and structure.

Avoid global variables

The program_name and start_time variables are global variables. It's generally better to explicitly pass variables your function will need rather than using the vague implicit linkage of a global variable. If they must be globals (which is not the case here), then make them static.

Be careful with size assumptions

The code currently has these three lines:

unsigned int lo,hi;
__asm__ __volatile__ ("rdtsc" : "=a" (lo), "=d" (hi));
return ((uint64_t)hi << 32) | lo;


It's not wrong as it is, but I'd probably write that using explicit uint32_t sizes for hi and lo to eliminate the possibility of a compiler with a 16-bit unsigned int.

Isolate code to a function

Right now, the timing is done in main. I'd probably rewrite that to contain the timing code in a function:

double approx_CPU_MHz(unsigned sleeptime) {
using namespace std::chrono_literals;
uint64_t cycles_start = get_cycles();
double time_start = age();
uint64_t elapsed_cycles = get_cycles() - cycles_start;
double elapsed_time = age() - time_start;
return elapsed_cycles / elapsed_time / 1000000.0;
}


Think of the user

The PrintUsage routine looks like this:

void PrintUsage()
{
printf("Usage:\n");
printf("%s [measurment duration]\n", program_name);
}


There are three problems with that from the user's point of view. First, it doesn't tell the user the unit of measure for the measurement duration. Is it seconds? Minutes? Microseconds? The user would have to look at the source code to answer that question. Second, it could reasonably be interpreted to be two arguments, both of which are optional. Putting an underscore instead of a space would fix that. Third, the word "measurement" is misspelled.

Use iostream

It's not wrong to use printf but using iostreams is both more C++-like and can actually save some runtime processing. To evaluate printf, the computer has to interpret the format string first, while using << means that the compiler has already evaluated the argument type at compile-time.