矩阵乘法中的共享内存涉及多个进程

Question

I am trying to learn IPC via shared memory . I don't understand one thing how do I allocate shared memory to various variables of different processes.For example , in a matrix multiplication i don't want to declare matrix globally but via shared memory . How do we attach memory to different matrices as :

Could anyone help me with this ? Also , can it be replaced by mmap() and shm_open() or is it just files ? If yes , how ? I could not find apt examples in which how shared memory is modified by the processes is illustrated .

Answer 1

First of all: Don't use shared memory for IPC. It's messy.

In most situations where shared memory would make sense, it's easier and more appropriate to just use threads (so that all memory is shared). When that isn't possible, you're probably better off using some sort of message-passing scheme.

How do I allocate memory from this shared memory for matrix A , B and resultant matrix and
attach to it as I don't have to declare matrices globally and each process would need ABC to do
part of problem.

You'd have to come up with some way to allocate memory within the shared memory segment. You can't use malloc() for this, since its arena doesn't include the shared memory.

If the only data you're going to store is going to be those three matrices, and the size of those matrices is known to both processes, you could just stack them all into a structure and store that in the segment:

struct shm_data {
    float A[123][123], B[123][123], C[123][123];
};

struct shm_data *mat = (struct shm_data *) shm;

If the size of the matrices isn't fixed, though, you'd have to figure their addresses at runtime.

Also , can it be replaced by mmap() and shm_open() or is it just files ?

Yes. In fact, mmap() is almost universally preferable to shm_open() , as the memory segment is backed by a file and obeys filesystem semantics, so it can be cleaned up sensibly. POSIX shared memory (which is what you're using) is much weirder.