Skip to content

Intra-node shared memory (SHM) optimizations for CPU primitives #458

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 17 commits into
base: main
Choose a base branch
from
+624 −2
Open

Intra-node shared memory (SHM) optimizations for CPU primitives #458

Show file tree
Hide file tree
Changes from 13 commits
Commits
Show all changes
17 commits
Select commit Hold shift + click to select a range
1738e8e
add shm allreduce
gaopengff Jul 11, 2025
76d1114
add bf16 and half support
gaopengff Jul 16, 2025
2d152a3
remove bf16 support
gaopengff Jul 18, 2025
8c29eeb
add bf16 support
gaopengff Jul 22, 2025
554d317
use reduce function to do reduce job
gaopengff Jul 24, 2025
0fdde35
refine format
gaopengff Jul 24, 2025
be7da7c
fix accuracy issue
gaopengff Aug 11, 2025
5b698dc
move intro-node check to allreduce()
gaopengff Aug 18, 2025
3564e95
remove debug code
gaopengff Aug 20, 2025
3ac8065
use local_rank to check intra-node condition
gaopengff Aug 25, 2025
8d3ae22
add support for multi-thread and code for local reduction
gaopengff Aug 29, 2025
2985c3e
move intra-node check to createAndConnectAllPairs
gaopengff Sep 1, 2025
704d35d
update to main
gaopengff Sep 1, 2025
9ae2842
add check for gpu input and fix format issue
gaopengff Sep 3, 2025
1b7660a
fix format issue
gaopengff Sep 3, 2025
ab9c63b
fix timeout ut
gaopengff Sep 5, 2025
9f726e9
add fininalize method and fix ut
gaopengff Sep 12, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions gloo/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,6 +51,11 @@ list(APPEND GLOO_HDRS
"${CMAKE_CURRENT_SOURCE_DIR}/types.h"
)

if(NOT MSVC)
list(APPEND GLOO_SRCS "${CMAKE_CURRENT_SOURCE_DIR}/allreduce_shm.cc")
list(APPEND GLOO_HDRS "${CMAKE_CURRENT_SOURCE_DIR}/allreduce_shm.h")
endif()

if(USE_CUDA)
file(GLOB GLOO_CUDA_SRCS
"${CMAKE_CURRENT_SOURCE_DIR}/cuda*.cc"
Expand Down
15 changes: 14 additions & 1 deletion gloo/allreduce.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
#include "gloo/common/logging.h"
#include "gloo/math.h"
#include "gloo/types.h"
#include "gloo/allreduce_shm.h"

namespace gloo {

Expand Down Expand Up @@ -131,14 +132,25 @@ void allreduce(const detail::AllreduceOptionsImpl& opts) {
return;
}

switch (opts.algorithm) {
auto algorithm = opts.algorithm;

#ifndef _WIN32
if (context->isIntraNode()) {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we add a check to make sure inputs are on CPU before setting it to shm? I believe this would fail when using local ibv cuda backend

Can also check device->hasGPUDirect() and instead disable for all GPU direct supported backends

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure. I've changed the check condition to "(context->isIntraNode() && !context->getDevice()->hasGPUDirect())".

algorithm = detail::AllreduceOptionsImpl::SHM;
}
#endif

switch (algorithm) {
case detail::AllreduceOptionsImpl::UNSPECIFIED:
case detail::AllreduceOptionsImpl::RING:
ring(opts, reduceInputs, broadcastOutputs);
break;
case detail::AllreduceOptionsImpl::BCUBE:
bcube(opts, reduceInputs, broadcastOutputs);
break;
case detail::AllreduceOptionsImpl::SHM:
shm(opts);
break;
default:
GLOO_ENFORCE(false, "Algorithm not handled.");
}
Expand All @@ -153,6 +165,7 @@ void ring(
const auto slot = Slot::build(kAllreduceSlotPrefix, opts.tag);
const size_t totalBytes = opts.elements * opts.elementSize;


// Note: context->size > 1
const auto recvRank = (context->size + context->rank + 1) % context->size;
const auto sendRank = (context->size + context->rank - 1) % context->size;
Expand Down
1 change: 1 addition & 0 deletions gloo/allreduce.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,6 +39,7 @@ struct AllreduceOptionsImpl {
UNSPECIFIED = 0,
RING = 1,
BCUBE = 2,
SHM = 3,
};

explicit AllreduceOptionsImpl(const std::shared_ptr<Context>& context)
Expand Down
Loading
Loading