在Xeon E5 2650 v4上不能同时启动超过12个线程

[huyonghuyong]

大家好，最近买了一台联想至强E5 2650 v4服务器，就想把之前的多线程图像处理服务程序搬到至强上来试试，毕竟12核24线程。直接开了20个线程测试并发效果，结果大跌眼镜还没有i7运行效果好。开的20个线程只有12个线程同时跑起来了，其他核心一动没动。使用intel parallel studio分析，尝试了各种优化办法都没有找到原因。同时用Qt4.8写了一个模拟程序之前服务程序的多线程操作方式，在任务处理时间小于window时间片的情况下仍然不能同时将20个线程运行起来。demo代码如下

#include <QtCore/QCoreApplication>
#include <QThread>
#include <QSemaphore>
#include <QMutex>
#include <Windows.h>
#define THREAD_NUMBER	20
QSemaphore	m_producer;
QSemaphore	m_customer;
volatile long long GolbalSum = 0;
class MyThread: public QThread
{
public:
	MyThread(QObject *parent=0);
	~MyThread();
	void run();
public:
	long long	*m_sum;
	bool		m_bExit;
	QMutex		m_singleTask;
	QMutex		m_TaskDone;
};
MyThread::MyThread(QObject *parent/* =0 */):QThread(parent)
{
	m_sum				= 0;
	m_bExit				= false;
	m_TaskDone.lock();
	start();
}
MyThread::~MyThread()
{
	m_bExit				= true;
	m_singleTask.unlock();
}
void MyThread::run()
{
	m_singleTask.lock();
	m_singleTask.lock();
	while (!m_bExit)
	{
		if (!m_producer.tryAcquire(1, 1))
		{
			continue;
		}
		// handle the task
		for (int i = 0; i < 500000; ++i)
			(*m_sum)++;
		
		m_customer.release();
		m_singleTask.lock();
	}
}
int main(int argc, char *argv[])
{
	QCoreApplication a(argc, argv);
	int loopCnt = 1000;
	MyThread threads[THREAD_NUMBER];
	long long threadsSum[THREAD_NUMBER];
	LARGE_INTEGER freq_detect, before_detect, after_detect;
	QueryPerformanceFrequency(&freq_detect);
	while(1)
	{
		for (int i = 0; i<THREAD_NUMBER; ++i)
		{
			threads[i].m_sum = &threadsSum[i];
			threads[i].m_singleTask.unlock();
			
		}
		if (0)//loopCnt-- <= 0)
		{
			for (int i = 0; i<THREAD_NUMBER; ++i)
			{
				threads[i].m_bExit = true;
				threads[i].wait();
			}
			break;
		}
		Sleep(8);
		
		QueryPerformanceCounter(&before_detect);
		m_producer.release(THREAD_NUMBER);
		m_customer.acquire(THREAD_NUMBER);
		QueryPerformanceCounter(&after_detect);
		printf("%f %f %f\t", before_detect.QuadPart*1e3/freq_detect.QuadPart, 
			after_detect.QuadPart*1e3/freq_detect.QuadPart, (after_detect.QuadPart-before_detect.QuadPart)*1e3/freq_detect.QuadPart);
		for (int i=0; i<THREAD_NUMBER; ++i)
		{
			GolbalSum += *threads[i].m_sum;
		}
		printf("Current sum:%I64d\n", GolbalSum);
	}
	//return a.exec();
}

在使用setthreadaffinity()函数情况下，将线程绑定到固定核心上是能够将20个线程同时调度起来的，但是会受到系统中的其他进程的影响，处理时间不稳定。
不知道各位有没有对多线程并发比较熟悉的，希望能得到一点意见，感谢。