Skip to content
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

Jump to bottom

Adding comments #107

Merged
merged 3 commits into from
Aug 17, 2017
Merged

Adding comments #107

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
189d447
Remove unnecessary comment
iwiwi Aug 15, 2017
de0ff50
Describe why we don't use MPI4py's scatter
iwiwi Aug 15, 2017
9730427
Describe why we delay the communicator initialization
iwiwi Aug 15, 2017
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: 4 additions & 1 deletion chainermn/communicators/_base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,7 +96,10 @@ def __init__(self, mpi_comm, use_nccl):

self._init_ranks()

# TODO(akiba): write why we delay initializing comms
# We have to delay the initialization of communicators. This is because
# NCCL's communicators use the current CUDA devices at the time of
# initialization. Therefore, we have to initialize NCCL communicators
# after users set the devices to use.
self.inter_mpi_comm = None
self.intra_mpi_comm = None
if self.use_nccl:
Expand Down
2 changes: 0 additions & 2 deletions chainermn/communicators/hierarchical_communicator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,8 +38,6 @@ def allreduce_grad(self, model):
self.gpu_buffer_a.ptr(), self.gpu_buffer_b.ptr(), n_elems_total,
nccl.NCCL_FLOAT, nccl.NCCL_SUM, 0, stream.ptr)

# TODO(akiba): sync necessary?

# Inter-node allreduce
if self.intra_rank == 0:
_communication_utility.inter_allreduce_gpu(
Expand Down
3 changes: 2 additions & 1 deletion chainermn/dataset.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,8 @@ def scatter_dataset(dataset, comm):
assert hasattr(comm, 'send')
assert hasattr(comm, 'recv')

# TODO(akiba): write why we do not use mpi_comm.scatter
# We cannot use `mpi_comm.scatter`. This is due to MPI4py's bug.
# For large datasets, when using `mpi_comm.scatter`, it causes MemoryError.
if comm.rank == 0:
mine = None
n_total_samples = len(dataset)
Expand Down