Ahh, okay. I think the main confusion is I was kind of ignoring the code portion of the library, which is easily deduplicated like you said.

Any allocations done by the library running (opencl.so in this example) will still be separate per process though, unless the program is passing mutable shared memory buffers around, which could be the case if there’s IPC going on.

The point about memory usage being hard to measure is definitely true, especially since linux won’t actually consume memory until a write is made to the block, unlike Windows which preemptively allocates the full requested size.