Intra-node shared memory (SHM) optimizations for CPU primitives #458
Conversation
|
Hi @gaopengff! Thank you for your pull request and welcome to our community. Action RequiredIn order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with If you have received this in error or have any questions, please contact us at [email protected]. Thanks! |
gloo/allreduce.cc
Outdated
| if (is_intra_node(context->size)) { | ||
| shm(opts); | ||
| return; | ||
| } |
There was a problem hiding this comment.
can we move this into the allreduce() function above? fits a bit better there since we pick the algorithm at that level
may be nice to add a RING_LOCAL or RING_SHMEM algorithm selector as well
There was a problem hiding this comment.
Thanks for your advice. We have added a new algorihtm selectordetail::AllreduceOptionsImpl::SHM and moved statement of checking intra-node to allreduce().
gloo/allreduce.cc
Outdated
| @@ -153,6 +154,15 @@ void ring( | |||
| const auto slot = Slot::build(kAllreduceSlotPrefix, opts.tag); | |||
| const size_t totalBytes = opts.elements * opts.elementSize; | |||
|
|
|||
|
|
|||
| if (is_intra_node(context->size)) { | |||
There was a problem hiding this comment.
we should also check if the tensor is a CUDA tensor and bypass if it is
There was a problem hiding this comment.
For cuda check, I think it's not necessary here. If the input tensors' device is cuda, pytorch will either copy cuda tensors to cpu before calling gloo::allreduce or use gloo::CudaAllreduceRingChunked do reduce cuda input directly. In that case gloo::allreduce is only for cpu inputs. Refer to pytorch code at ProcessGroupGlooCuda.cpp
|
|
||
| bool is_intra_node(const int size) { | ||
| // must launch with torchrun | ||
| auto local_size_string = std::getenv("LOCAL_WORLD_SIZE"); |
There was a problem hiding this comment.
I don't think this check is safe -- for torchft for instance we often run with Gloo only cross host and if you're using an 8x8 configuration this would trigger shm logic for cross host comms
There was a problem hiding this comment.
In my understanding, world_size is 64 and the local_size is 8 if we launch program with 8x8 configuration, which won't introduce shm allreduce.
This PR is for RFC #455. It has implemented shm allreduce.