How to replicate a page from one GPU to another GPU?

[Pikaasa]

I have a page already available on GPU1, and I want to replicate(not migrate) the same page on GPU2. Which files should I check for this? Any pointers would be really helpful. Thanks!

[syifan]

Replicating a page on another GPU can be a somewhat complex process. Could you let me know how you imagine real hardware supporting this feature? Also, to what degree do you want to model this behavior?

The starting point should be the mmu.go. https://github.com/sarchlab/akita/blob/v4/mem/vm/mmu/mmu.go. You may want to modify the page table https://github.com/sarchlab/akita/blob/v4/mem/vm/pagetable.go to be able to record multiple owning GPUs. Or you can let the page table to have multiple entries for one virtual address. Then, you can let the MMU respond to different GPUs with their own physical addresses.

[Pikaasa]

I'm looking to implement the "Read Page Duplication" strategy from the HPCA 2024 GRIT paper using MGPUSIM. The idea is that when a GPU encounters a page fault on a read, it replicates the page in its local memory. If multiple GPUs read the same page, each keeps a replica. On a write, a protection fault triggers invalidation across GPUs, ensuring consistency before proceeding.

[Pikaasa]

Sir, Is there a way to mark a page as read-only when it is migrated in MGPUSim?

[syifan]

You can always add a field in page table to mark it as read-only.