Use more CPU cores

[satoshi]

Hi there,
is there a way to run a Qt application using all the cores of the CPU? I've a two-core CPU and there are two threads in my program. This application has to be as fast as possible, a smart thing will be assign a thread for each core. Is it possible?

If yes, is there a way to determine the number of CPU cores of the machine where the application is running?

Sorry for my english

Thanks, Dario

[wysota]

If there are two threads in your application then if your system policy allows it, two cores will be used.

[satoshi]

If there are two threads in your application then if your system policy allows it, two cores will be used.

Thanks, this is what I was expecting but it seems it's using only one core.

I'm running Qt 4.5.3 under Windows XP. The effect is the same under Ubuntu.

Any suggestion?

[wysota]

Maybe you are not using two threads?

[satoshi]

Maybe you are not using two threads?

I see two threads when I debug...

[wysota]

I see two threads when I debug...

Which doesn't mean both are doing something. If you have bugs in your code concerning thread affinity, then one of the threads will be idle all the time and won't use practically any cpu power.

[satoshi]

Which doesn't mean both are doing something. If you have bugs in your code concerning thread affinity, then one of the threads will be idle all the time and won't use practically any cpu power.

Ok, thanks. I'll verify that.

Dario

[satoshi]

Maybe I'm not using threads correctly.. Here's my code:

for(int i = 0; i < numCores; i++)
{
        new MyThread(...);
}

for(int i = 0; i < numCores; i++)
{
        new MyThread(...);
}

To copy to clipboard, switch view to plain text mode 

The class MyThread is:

//HEADER
#ifndef MYTHREAD_H
#define MYTHREAD_H
 
#include <QThread>
 
class MyThread : public QThread
{
    public:
        MyThread(...);
 
    protected:
        void run();
};
 
#endif // MYTHREAD_H
 
 
// BODY
#include "mythread.h"
 
MyThread::MyThread(...)
{
    start();
}
 
void MyThread::run()
{
    ...
}

//HEADER
#ifndef MYTHREAD_H
#define MYTHREAD_H

#include <QThread>

class MyThread : public QThread
{
    public:
        MyThread(...);

    protected:
        void run();
};

#endif // MYTHREAD_H


// BODY
#include "mythread.h"

MyThread::MyThread(...)
{
    start();
}

void MyThread::run()
{
    ...
}

To copy to clipboard, switch view to plain text mode 

Is it right? Have I to call exec() in the run() method? If yes, where?

Thanks,

Dario

[wysota]

start() should not be called from the constructor, that's for sure. But what is important is the contents of the run() method too.

[satoshi]

Ok, I've removed the start() method in the constructor of MyThread, now the code is:

MyThread *mt;
for(int i = 0; i < numCores; i++)
{
        mt = new MyThread(...);
	mt->start();
}

MyThread *mt;
for(int i = 0; i < numCores; i++)
{
        mt = new MyThread(...);
	mt->start();
}

To copy to clipboard, switch view to plain text mode 

The result is the same, it will only create one new thread (numCores = 2)...

The content of run() isn't nothing special, say for(int i = 0; i < INT_MAX; i++){}

Thanks for the help you're giving me.

Dario

[satoshi]

I think the problem is the debugger, since if i put numCores = 50, the debugger says that 10 threads has been created. This is too strange to be real, or not?

Anyway, always one core is 100% busy, the second is free.

[satoshi]

Ok, now I'm sure that the threads are running correctly. With numCores=2, if I put a breakpoint in the execution of the run() method, the debugger will stop there sometimes with one thread (id=4) and sometimes with the other thread (id=5).

Always one core is 100% busy, the second is free.

[wysota]

Open the task manager, right click on your process and check its processor affinity. Both checkboxes should be checked.

[borisbn]

1. Use QThread::idealThreadCount to retrieve processors core count
2. I think, that one of your thread doesn't give any time to another. Try to call

msleep( 1 );

msleep( 1 );

To copy to clipboard, switch view to plain text mode 

inside for in your run functions as following:

for(int i = 0; i < INT_MAX; i++)
{
    if ( ( i % 100 ) == 0 )
    {
        msleep( 1 );
    }
}

for(int i = 0; i < INT_MAX; i++)
{
    if ( ( i % 100 ) == 0 )
    {
        msleep( 1 );
    }
}

To copy to clipboard, switch view to plain text mode 

[satoshi]

Open the task manager, right click on your process and check its processor affinity. Both checkboxes should be checked.

Both checked

1. Use QThread::idealThreadCount to retrieve processors core count
2. I think, that one of your thread doesn't give any time to another. Try to call

msleep( 1 );

msleep( 1 );

To copy to clipboard, switch view to plain text mode 

inside for in your run functions as following:

for(int i = 0; i < INT_MAX; i++)
{
    if ( ( i % 100 ) == 0 )
    {
        msleep( 1 );
    }
}

for(int i = 0; i < INT_MAX; i++)
{
    if ( ( i % 100 ) == 0 )
    {
        msleep( 1 );
    }
}

To copy to clipboard, switch view to plain text mode 

Thanks for that but sorry, I just discovered that the threads are running correctly...


So, after starting this new 2 thread, I've 4 threads running in the application.

The first thread (which created the second) is waiting for a timer which will timeout in 3 minutes. This is a QApplication.
The second thread (which created the 2 new threads) is waiting for the termination of these new 2 threads (with QWaitCondition), which are calculating and need a lot of CPU.

I'm wondering if the OS is stupid and put the first 2 threads in one core and the others 2 in the other thread...

[wysota]

The affinity should be dynamically changing to balance the load between two cores. If one thread waits on a wait condition, it is not scheduled by the OS so it won't add to the overall load of the processing unit. Maybe it is your OS that reports usage incorrectly.

[spud]

start() should not be called from the constructor, that's for sure.

Why not? That way the user of the class cannot forget to call it.

[wysota]

For one thing start() calls a method that calls a virtual method. It's true that the virtual method is called on behalf of another thread but I won't guarantee the object will be fully constructed by then. So it might work in 99% of the cases but it will fail when you need it the most (acc. to Murphy's Law).

[spud]

Ok, fair enough. I can't see that start() calls any virtual methods in Qt 4.6, though.

P.S. I'm not nitpicking, it's just that if it's not safe, I'll have too change a lot of code.

[wysota]

Ok, fair enough. I can't see that start() calls any virtual methods in Qt 4.6, though.

It calls QThreadPrivate::start() which in turn calls QThread::run() (which is virtual).

P.S. I'm not nitpicking, it's just that if it's not safe, I'll have too change a lot of code.

As I said, in 99% of the cases it will work without problems. But if start subclassing and adding variables, something might explode.