How to bind a thread to specific thread in multithread application?

Dear all;

I am going to use "pthread_setaffinity_np" to bind a thread to a specific core. My application has two threads. I have used mutex to assign a specific id to each thread and then bind that thread to a core different from another core.
but it seems that the os assigns both thread to one core.
What should I do to bind each thread to a specific core?

id = 0;
...
cpu_set_t myset;
pthread_mutex_lock(&mutex);
MyNum = id;
CPU_ZERO(&myset);
CPU_SET(MyNum+1, &myset);
pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t),&myset);
id ++;
pthread_mutex_unlock(&mutex);

...

Thanks in advance for your help.

If you talk about core I think you have a multicore CPU. So "pthread_setaffinity_np" function can not do anything useful for you, because you only have one CPU (with more cores), if I did read it right at http://www.unix.com/man-page/All/3/P...ETAFFINITY_NP/.
I think there is a discussion at GCC, glibc about that problem, too.
I also think it is easier to read out the core's register (SYSCFG (a MSR) I guess) and exit the pthread on a control mechanism.

Maby I can help you to find a way out if you explain, why you need to have a thread running on a special core.

Hi Bastl,

Thank you very much for your help.
I have run a multithread application on ubuntu. While I watch the system monitor, just one cpu is completely used, while the other one has almost zero cpu usage.
This property is against the goal of multithread applications. I think if the os distributes each thread on a core, it will gain better performance.

In System monitor, it shows two cpus for my system. What is the difference between core and cpu then?

O.K., I don't know ubuntu verry much but the thread optimization is done by the kernel.
You need a kernel with multicore support enabled ( Multi-core scheduler support (SCHED_MC) ).
Then on a multicore CPU ( 4 or more cores ) every time a thread/process is created the thread/process is set to the next core.
Duo core machines are handled different but I can not say if they are handled the way like in your case - wouldn't make much sense, I have to agree, - but they are limited.
There are multiprocessor systems where the system resouces have to (can) be set up for each processor.
A IA-Multicore processor it verry limited about that (pins, external connections).
Else there are no differences (compute time).
On the other hand I don't know what that monitor app. is representing (does measuring).

Originally Posted by h.aliee id = 0; ... cpu_set_t myset; pthread_mutex_lock(&mutex); MyNum = id; CPU_ZERO(&myset); CPU_SET(MyNum+1, &myset); pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t),&myset); id ++; pthread_mutex_unlock(&mutex); ...

Hey,

I'm not sure if that solves your problems, but the CPU number usually starts at 0 and NOT 1 (you could easily verify this by using top(1) or the sysFS). So, maybe you could try CPU_SET(MyNum, &myset) instead?

Secondly, please check and verify the return value of 'pthread_setaffinity_np'. It should tell you if you specified an incorrect CPU mask.

I hope that helps,
Andi