512
votes

Is there a way to determine how many cores a machine has from C/C++ in a platform-independent way? If no such thing exists, what about determining it per-platform (Windows/*nix/Mac)?

19
If you want to use it find out how many threads to start please use NUMBER_OF_PROCESSORS as primary measure. I leave it as an excercise to you why this is much better (if people would use it more) then using hardware cores. How much cores belong to your program are an environmental issue!Lothar
Note that std::thread::hardware_concurrency returns the number of physical CPU cores, but nproc in Linux shows only the number of CPU cores the current process can run on, which can be controlled with sched_setaffinity. I haven't found a way to get that instead from standard C++:, see e.g. in Python: stackoverflow.com/questions/1006289/…Ciro Santilli 新疆再教育营六四事件法轮功郝海东

19 Answers

770
votes

C++11

#include <thread>

//may return 0 when not able to detect
const auto processor_count = std::thread::hardware_concurrency();

Reference: std::thread::hardware_concurrency


In C++ prior to C++11, there's no portable way. Instead, you'll need to use one or more of the following methods (guarded by appropriate #ifdef lines):

  • Win32

    SYSTEM_INFO sysinfo;
    GetSystemInfo(&sysinfo);
    int numCPU = sysinfo.dwNumberOfProcessors;
    
  • Linux, Solaris, AIX and Mac OS X >=10.4 (i.e. Tiger onwards)

    int numCPU = sysconf(_SC_NPROCESSORS_ONLN);
    
  • FreeBSD, MacOS X, NetBSD, OpenBSD, etc.

    int mib[4];
    int numCPU;
    std::size_t len = sizeof(numCPU); 
    
    /* set the mib for hw.ncpu */
    mib[0] = CTL_HW;
    mib[1] = HW_AVAILCPU;  // alternatively, try HW_NCPU;
    
    /* get the number of CPUs from the system */
    sysctl(mib, 2, &numCPU, &len, NULL, 0);
    
    if (numCPU < 1) 
    {
        mib[1] = HW_NCPU;
        sysctl(mib, 2, &numCPU, &len, NULL, 0);
        if (numCPU < 1)
            numCPU = 1;
    }
    
  • HPUX

    int numCPU = mpctl(MPC_GETNUMSPUS, NULL, NULL);
    
  • IRIX

    int numCPU = sysconf(_SC_NPROC_ONLN);
    
  • Objective-C (Mac OS X >=10.5 or iOS)

    NSUInteger a = [[NSProcessInfo processInfo] processorCount];
    NSUInteger b = [[NSProcessInfo processInfo] activeProcessorCount];
    
214
votes

This functionality is part of the C++11 standard.

#include <thread>

unsigned int nthreads = std::thread::hardware_concurrency();

For older compilers, you can use the Boost.Thread library.

#include <boost/thread.hpp>

unsigned int nthreads = boost::thread::hardware_concurrency();

In either case, hardware_concurrency() returns the number of threads that the hardware is capable of executing concurrently based on the number of CPU cores and hyper-threading units.

65
votes

OpenMP is supported on many platforms (including Visual Studio 2005) and it offers a

int omp_get_num_procs();

function that returns the number of processors/cores available at the time of call.

39
votes

If you have assembly-language access, you can use the CPUID instruction to get all sorts of information about the CPU. It's portable between operating systems, though you'll need to use manufacturer-specific information to determine how to find the number of cores. Here's a document that describes how to do it on Intel chips, and page 11 of this one describes the AMD specification.

35
votes

(Almost) Platform Independent function in c-code

#ifdef _WIN32
#include <windows.h>
#elif MACOS
#include <sys/param.h>
#include <sys/sysctl.h>
#else
#include <unistd.h>
#endif

int getNumCores() {
#ifdef WIN32
    SYSTEM_INFO sysinfo;
    GetSystemInfo(&sysinfo);
    return sysinfo.dwNumberOfProcessors;
#elif MACOS
    int nm[2];
    size_t len = 4;
    uint32_t count;

    nm[0] = CTL_HW; nm[1] = HW_AVAILCPU;
    sysctl(nm, 2, &count, &len, NULL, 0);

    if(count < 1) {
        nm[1] = HW_NCPU;
        sysctl(nm, 2, &count, &len, NULL, 0);
        if(count < 1) { count = 1; }
    }
    return count;
#else
    return sysconf(_SC_NPROCESSORS_ONLN);
#endif
}
17
votes

On Linux, you can read the /proc/cpuinfo file and count the cores.

11
votes

Note that "number of cores" might not be a particularly useful number, you might have to qualify it a bit more. How do you want to count multi-threaded CPUs such as Intel HT, IBM Power5 and Power6, and most famously, Sun's Niagara/UltraSparc T1 and T2? Or even more interesting, the MIPS 1004k with its two levels of hardware threading (supervisor AND user-level)... Not to mention what happens when you move into hypervisor-supported systems where the hardware might have tens of CPUs but your particular OS only sees a few.

The best you can hope for is to tell the number of logical processing units that you have in your local OS partition. Forget about seeing the true machine unless you are a hypervisor. The only exception to this rule today is in x86 land, but the end of non-virtual machines is coming fast...

8
votes

One more Windows recipe: use system-wide environment variable NUMBER_OF_PROCESSORS:

printf("%d\n", atoi(getenv("NUMBER_OF_PROCESSORS")));
7
votes

You probably won't be able to get it in a platform independent way. Windows you get number of processors.

Win32 System Information

6
votes

Windows (x64 and Win32) and C++11

The number of groups of logical processors sharing a single processor core. (Using GetLogicalProcessorInformationEx, see GetLogicalProcessorInformation as well)

size_t NumberOfPhysicalCores() noexcept {

    DWORD length = 0;
    const BOOL result_first = GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &length);
    assert(GetLastError() == ERROR_INSUFFICIENT_BUFFER);

    std::unique_ptr< uint8_t[] > buffer(new uint8_t[length]);
    const PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX info = 
            reinterpret_cast< PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX >(buffer.get());

    const BOOL result_second = GetLogicalProcessorInformationEx(RelationProcessorCore, info, &length);
    assert(result_second != FALSE);

    size_t nb_physical_cores = 0;
    size_t offset = 0;
    do {
        const PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX current_info =
            reinterpret_cast< PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX >(buffer.get() + offset);
        offset += current_info->Size;
        ++nb_physical_cores;
    } while (offset < length);
        
    return nb_physical_cores;
}

Note that the implementation of NumberOfPhysicalCores is IMHO far from trivial (i.e. "use GetLogicalProcessorInformation or GetLogicalProcessorInformationEx"). Instead it is rather subtle if one reads the documentation (explicitly present for GetLogicalProcessorInformation and implicitly present for GetLogicalProcessorInformationEx) at MSDN.

The number of logical processors. (Using GetSystemInfo)

size_t NumberOfSystemCores() noexcept {
    SYSTEM_INFO system_info;
    ZeroMemory(&system_info, sizeof(system_info));
    
    GetSystemInfo(&system_info);
    
    return static_cast< size_t >(system_info.dwNumberOfProcessors);
}

Note that both methods can easily be converted to C/C++98/C++03.

5
votes

More on OS X: sysconf(_SC_NPROCESSORS_ONLN) is available only versions >= 10.5, not 10.4.

An alternative is the HW_AVAILCPU/sysctl() BSD code which is available on versions >= 10.2.

4
votes

Windows Server 2003 and later lets you leverage the GetLogicalProcessorInformation function

http://msdn.microsoft.com/en-us/library/ms683194.aspx

4
votes

Unrelated to C++, but on Linux I usually do:

grep processor /proc/cpuinfo | wc -l

Handy for scripting languages like bash/perl/python/ruby.

3
votes

hwloc (http://www.open-mpi.org/projects/hwloc/) is worth looking at. Though requires another library integration into your code but it can provide all the information about your processor (number of cores, the topology, etc.)

3
votes

On linux the best programmatic way as far as I know is to use

sysconf(_SC_NPROCESSORS_CONF)

or

sysconf(_SC_NPROCESSORS_ONLN)

These aren't standard, but are in my man page for Linux.

3
votes

On Linux, it's may not be safe to to use _SC_NPROCESSORS_ONLN as it's not part of POSIX standard and the sysconf manual states as much. So there's a possibility that _SC_NPROCESSORS_ONLN may not be present:

 These values also exist, but may not be standard.

     [...]     

     - _SC_NPROCESSORS_CONF
              The number of processors configured.   
     - _SC_NPROCESSORS_ONLN
              The number of processors currently online (available).

A simple approach would be to read /proc/stat or /proc/cpuinfo and count them:

#include<unistd.h>
#include<stdio.h>

int main(void)
{
char str[256];
int procCount = -1; // to offset for the first entry
FILE *fp;

if( (fp = fopen("/proc/stat", "r")) )
{
  while(fgets(str, sizeof str, fp))
  if( !memcmp(str, "cpu", 3) ) procCount++;
}

if ( procCount == -1) 
{ 
printf("Unable to get proc count. Defaulting to 2");
procCount=2;
}

printf("Proc Count:%d\n", procCount);
return 0;
}

Using /proc/cpuinfo:

#include<unistd.h>
#include<stdio.h>

int main(void)
{
char str[256];
int procCount = 0;
FILE *fp;

if( (fp = fopen("/proc/cpuinfo", "r")) )
{
  while(fgets(str, sizeof str, fp))
  if( !memcmp(str, "processor", 9) ) procCount++;
}

if ( !procCount ) 
{ 
printf("Unable to get proc count. Defaulting to 2");
procCount=2;
}

printf("Proc Count:%d\n", procCount);
return 0;
}

The same approach in shell using grep:

grep -c ^processor /proc/cpuinfo

Or

grep -c ^cpu /proc/stat # subtract 1 from the result
2
votes

OS X alternative: The solution described earlier based on [[NSProcessInfo processInfo] processorCount] is only available on OS X 10.5.0, according to the docs. For earlier versions of OS X, use the Carbon function MPProcessors().

If you're a Cocoa programmer, don't be freaked out by the fact that this is Carbon. You just need to need to add the Carbon framework to your Xcode project and MPProcessors() will be available.

2
votes

For Win32:

While GetSystemInfo() gets you the number of logical processors, use GetLogicalProcessorInformationEx() to get the number of physical processors.

-2
votes

you can use WMI in .net too but you're then dependent on the wmi service running etc. Sometimes it works locally, but then fail when the same code is run on servers. I believe that's a namespace issue, related to the "names" whose values you're reading.