Fix#209: improve socket transport performance
Split transfers over multiple sockets
Launch multiple threads to drive sockets
Detect AWS NICs and set nsockets/nthreads accordingly
Makes a change to allow the user to select between the static CUDA
runtime library (default) and the dynamic CUDA runtime library. Does
this by allowing `CUDARTLIB` to be overridden.
Added detection of IBM/Power NVLink bridge device.
Add NUMA support to PCI distance calculations.
Added NCCL_IGNORE_CPU_AFFINITY env var.
Fix memory leaks; GithubIssue#180
Compiler warning fix; GithubIssue#178
Replace non-standard variable length arrays. GithubIssue#171
Fix Tree+Shared Memory crash. GithubPR#185
Fix LL cleanup hang during long running DL jobs.
Fix NCCL_RINGS environment variable handling.
Added extra checks to catch repeat calls to ncclCommDestroy() GithubIssue#191
Improve bootstrap socket connection reliability at scale.
Fix hostname hashing issue. GithubIssue#187
Code cleanup to rename all non device files from *.cu to *.cc
The shared memory filename was only based on the destination. While
this was OK for rings since only one rank would send data to a given
rank, it would crash with trees because they communicate in both
directions.
Co-authored-by: Rong Ou <rong.ou@gmail.com>
Add tree algorithms for allreduce to improve performance at scale.
Add ncclCommAbort() and ncclCommGetAsyncError() to properly handle
network errors and be permit recover.
Detect initial CPU affinity and no longer escape it.
This is to handle EINTR/EGAIN properly (issue #137), and also
make the code consistent with the rest.
Unfortunately posix_fallocate and mmap do not follow the classic
return code/errno pattern, so we need to write wrappers around those
functions.
SYSCHECKVAL was not retrying when a retry was needed. Since not all
calls are inside a loop, that means we could silently miss an
EINTR/EAGAIN return code.
Also rework the socket connection code and improve error reporting.
Dynamically load external network from libnccl-net.so.
Add init function in networks.
Move PCI scoring to net.cu, only ask transport to provide a path.
Simplify CUDA PCI path detection.
Add dummy external network
Improved LL tuning for multi-node jobs.
Improved bootstrap for large job scaling.
Fixed a hang during bootstrap due to socket reuse.
Added operation name to the COLL INFO logging.
Add support for inter-node communication using sockets and InfiniBand/RoCE.
Improve latency.
Add support for aggregation.
Improve LL/regular tuning.
Remove tests as those are now at github.com/nvidia/nccl-tests .
Wenkai Du
Ke Wen
Hirochika Asai
Felix Abecassis
David Addison
jakirkham
Gustavo Alvarez
Rong Ou
Sylvain Jeaugey
Christian Sigg
Christian Sigg
Alex Sergeev
Ilya Biryukov
Nathan Luehr
Chad Whipkey