[highd]

Yes! I'm thinking of building something like this for my neural net training (1-2 days on AWS, 16 GPUs/processes on the job). In this case the "state" that I'd like to access is all the parameters of the model and training history, so I'm thinking I'll probably store an mmapped file so I can use other processes to poke at it while it's running. That way I can decouple the write-test-debug loops for the training code and the viz code.

[dllthomas]

> I'm thinking I'll probably store an mmapped file so I can use other processes to poke at it while it's running.

That seems to run substantial risk of seeing it in an inconsistent state, yeah?

[highd]

I generally use a semaphore when I'm reading and writing from my shm'd things. The data structure will also likely be append-only for the training process, as I want to see how things are changing over time.

Also I meant shm'd, not mmap'd.

[terminalcommand]

I am knew to the shared memory concept. I am familiar with named pipes. Could you please elaborate a bit, I'm curios.

Are you passing the reference to an mmap adress, or using the shm systemcalls? In what language are you programming in? Does race conditions endanger the shared memory? If so, how does using semaphores help?

Sorry, if I asked a lot of questions, feel free to answer any/none of them :)

[highd]

Sure! SHM is really cool, I just found out about it. It's an old Posix functionality, so people should use it more!

I'm using shm system calls in Python. Basically I get a buffer of raw bytes of a fixed size that is referred to by a key. When I have multiple processes running I just have to pass that key between them and they get access to that buffer of bytes to read and write.

On each iteration first I wait until the semaphore is free and then I lock it (P). That prevents anyone else from accessing the shared memory. I have the process read from the shared memory a set of variables - I have little helper functions that serialize and deserialize numpy arrays into raw bytes using fixed shapes and dtypes. Those arrays are then updated using some function combining the output of the process and the current value of the array. Then those arrays are reserialized and written back to the shm buffer as raw bytes again. Finally, the process releases the semaphore using P() so other processes can access it. The purpose of the semaphore is to prevent reading the arrays while another process is writing them - otherwise you might get interleaved old and new data from a given update. In a process-wise sense there is a race-condition, as each process can update at different times or in a different order, but for my purposes this is acceptable since neural net training is a stochastic sort of thing and it shouldn't care too much.

[0] http://nikitathespider.com/python/shm/ - original library which works fine for me

[1] http://semanchuk.com/philip/PythonIpc/ - updated version