High Performance Computing

Hi all,

(Hope this is the right forum for this question)

I have some multi-threaded C code (compiled with GCC 4.4.3) which accesses shared variables. Although I've marked those variables with volatile to guard against compiler re-ordering, I'm concerned that processor out-of-order execution may cause my code to fail, and I'm looking for a "low-cost" method of guaranteeing ordering is maintained in my code.

For example, I have something like...

Code:

memset(&task, 0, sizeof(task_t));/* null memory */
task.id.prefix = prefix_id;
task.id.instance = instance_id;
/* write-memorybarrier required here */
task.state = task_ready;

Where I need to ensure that the "task state" is only set to "task_ready" after the previous instructions have been committed. As the "task" is shared between threads, another thread seeing the state as "ready" may try to access its member variables, so it's vital that the tasks "prefix" and "instance" have been updated.

I know this is a common problem and mutexes and semaphores provide in-built memory barriers to address this problem but I'm trying to build a scalable application and I want to avoid their use if possible. I also know GCC provides built-in atomic operations but I see they involve locking the data-bus, and I've heard about system primitives like "smp_wmb()" but I'm not sure how to incorporate these into my "user-space" program as they are platform dependent.

Therefore can anyone provide pointers or advise on how best (in terms of scalability and speed) to guarantee ordering is maintained?

Thanks.