* for multinode gfx950, extend AR LL128 up to 256MB, extend RS LL128 up to 8MB per rank, extend AG LL up to 64KB per rank
* dont override direct allgather threshold if set to -1
* restore 2-node AR simple at earlier message sizes than higher multi-node AR
* extend range of LL for single-node RS on gfx950
* update algo/proto for multi-node allreduce on gfx942
* set single-node AR on gfx950 to Tree LL for KB message sizes
* decrease threshold for single node Tree for gfx950 AR
* Add initial commit to increase tb size to 512
* Fix LL perf issue when subset of NCCL_MAX_NTHREADS is used
Adding a constant to barrier_generic logic from using fallback logic when nthreads < NCCL_MAX_NTHREADS and nthreads == blockDim.X
* Adjust nthreads for LL
* Opt threads for reduce_scatter upper small range
* Add macro for single node
* Restrict MSCCL to 256 threads to prevent mem access fault
* Support pre-MI350 compatibility
* Partially refactor threadblock size override
* Use const macros instead of numerals
* opt out of unused function
* Batch P2P operations (2 per CU/channel) and update channel-part mapping
- Revert bitreversal and fix channel mapping to be compatible with P2P batching and avoid hangs
- P2P batching is only used for more than 2 nodes to avoid aggregating intra-node traffic when it is dominant for less than 2 nodes
* Address single node regression and channel per net peer
* Add batching threshold
* Add enable switch for batching
* Update CHANGELOG.md
* Add minor comment change
* Update CHANGELOG.md
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
* Update CHANGELOG.md
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
* change gfx950 algo/proto selection for multinode allreduce, allgather, reduceScatter
* gfx950 tuning: enable tuning for broadcast, allreduce starts LL128 earlier and switches to ring earlier, change LL128 start for allgather and reduceScatter
* lower LL128 threshold
* update reduceScatter LL128 min to match LL max for consistency
* enable multinode PXN and increase chunksize for gfx950
* change LL128 start to 128KB, adjust ring-start according to node-count
* disable code-path for fused-AR on LL128 for gfx950
* use LL128 starting from 1KB for multinode allgather on gfx950
* start LL128 earlier for multinode reduceScatter on gfx950
* start LL128 earlier for multinode broadcast on gfx950
* set multinode allreduce to start simple on 64MB for gfx950
* start LL128 from 1KB for multinode broadcast on gfx950
* setting multinode AR to use tree instead of ring at 16MB, 64MB, 128MB
* set multinode broadcast to use LL for up to 256KB depending on node-count for gfx950
* adjust algo for 32MB multinode allreduce on gfx950
* make 32MB tree LL128 for multinode AR on gfx950
* make sure ring is not picked on 2N allreduce on small sizes
Improve support for DirectNIC (CX8)
* Add support for XDR speed detection.
* When DirectNIC is enabled, report only the RDMA interfaces.
Extend the P2C (PXN over C2C) support to send/receive operations.
Support compilation with GCC 14 (Issues #1743, #1751).
Fix the unloading of network plugins that also provide tuner capability.
Fix the change of the current device across the calls to ncclCommDestroy()
and ncclCommAbort().
A note for users on MNNVL systems: please ensure an adequate stack size for
NCCL threads. While the default Linux stack size limit of 8192 KB is known
to be sufficient, we've seen crashes if the limit is changed to
"unlimited", as it causes the glibc library to unexpectedly *decrease* the
stack size of NCCL's background threads to just 2048 KB. Use "ulimit -s"
in bash to print the current limit; if needed, reset it to 8192 KB using
"ulimit -s 8192" (one also needs to ensure that the new setting is
propagated to other nodes when launching a multi-node NCCL job).
Improvements for GB200 systems
* Optimize the network performance by alternating the direction of the
rings and the NIC to GPU assignment across communicators to limit
unnecessary sharing.
* Fix the detection of C2C links in case GPU Direct RDMA is disabled
between a GPU and a NIC.
* Fix PXN support on MNNVL systems, where NCCL would try (and fail) to
share regular host memory across multiple nodes.
* Fix P2C (PXN over C2C), which is now preferred over regular PXN. This
support is currently preliminary and is disabled by default; use
NCCL_PXN_C2C=1 to enable.
Further reduce the overheads of CUDA graph capturing, which increased in
NCCL 2.26.2 for large graphs.
Optimize the network performance on DGX B200 systems by adjusting the
bandwidths provided to the graph search algorithm.
Enable fp8 reductions in symmetric kernels on Blackwell with CUDA 12.8.
Restore the plugin name handling logic to make it possible to specify a
path to the plugin (Issue #1732).
Restore the ability to change NCCL_COLLNET_ENABLE during execution
(Issue #1741).
Add an example tuner plugin with CSV-based overrides.
Remove an x86 dependency from the example profiler.
* Reapply "[AG and RS channel tuning] Add thread work threshold to tuning models and precompute reg index in LL128 (#1641)"
This reverts commit 943ad6f7820739385a0b54e81f823d0df1dbf71c.
* Decreasing NCCL_LL128_SHMEM_ELEMS_PER_THREAD from 16 to 8
Symmetric memory API and symmetric kernels
* Redesign from the ground up, enabling major latency and bandwidth
improvements.
* Add new API calls to register user-allocated memory among communicator
ranks into a NCCL window: ncclCommWindowRegister() and
ncclCommWindowDeregister(). The calls currently support symmetric
registration for P2P and NVLS, and require VMM memory buffers (i.e.,
CUMEM must be operational).
* Implement specialized kernels taking advantage of symmetrically
registered memory, with performance gains expected particularly for
small to medium message sizes.
* The kernels support 32 bit floating point types and smaller, and sum as
the reduction operator, with no more than one collective operation per
group.
* Floating point summation is always done in fp32 accumulators (with the
exception of fp8 on NVLS, where it uses fp16 inside the switch). Thus,
the accuracy with fp8 and fp16 data types should be much improved.
* This initial implementation supports non-network communicators only (P2P
and NVLS transports).
* To explore this functionality users need to use the new memory
registration API calls with the NCCL_WIN_COLL_SYMMETRIC flag and all
ranks of a communicator must pass buffers at the same offset in the same
registration when invoking a collective NCCL operation.
Add support for DGX Spark.
Add support for DirectNIC (CX8) to the internal IB plugin.
Add a new ncclCommShrink() API call
* It is a non-collective call similar to ncclCommSplit(), which makes it
possible to exclude some (possibly unresponsive) ranks from the parent
communicator.
Add support for loading multiple network plugins
* This enables the creation of generic containers that can work across a
range of providers.
* Allow NCCL_NET_PLUGIN to accept a comma-separated list of plugins to
load.
NVLink SHARP (NVLS) improvements
* Implement NVLS+IB SHARP support for AllGather and ReduceScatter with
user buffer registration. This improves performance and reduces the
number of CTAs needed to achieve peak bandwidth.
* Gracefully fall back by default to other transports if NVLS
initialization fails (the old behavior of returning an error code from a
NCCL call can be preserved by setting NCCL_NVLS_ENABLE=1).
* Decrease the NVLS channel count to 24 on Blackwell systems with multiple
NVLink domains per communicator.
* Enable fine-tuning of NCCL behavior per communicator using new
"ncclConfig_t" members "collnetEnable", "CTAPolicy", and "nvlsCTAs".
Profiler improvements
* Extend the init function by adding communicator name, comm id (hash),
rank, number of ranks, number of nodes, and the NCCL log function to the
argument list. This makes the name and the comm id available to all
events in the communicator without explicitly passing them to each
individual event. Add the communicator id and rank to the profiler trace
filename. Now, the communicator name can be set via a new "ncclConfig_t"
member "commName".
* Improve the accuracy of the GPU kernel events by providing GPU-generated
timestamps for the start and stop of every NCCL operation.
* Harmonize proxy events, removing overlaps between ProxyOp and ProxyStep
states.
* Add support for network-defined event updates (through
"recordEventState").
* Report the correct number of channels used by every collective/p2p
operation (used to be set to nMaxChannels for collectives and absent for
p2ps).
* Fix the logic on proxyCtrl Idle/Active events (Issue #1162).
* Fix an issue where the network proxy profiler could lose track of an
event identifier (Issue #1682).
* Improve the backward compatibility with plugins older than v4.
* Ensure that the work counters are 0-initialized.
* Fix a potential race condition in the network profiler that could result
in an event being linked to a wrong parent.
MNNVL improvements
* Increase to 16 the number of NICs used to communicate between MNNVL
domains on GB200 systems, to optimize the performance of collective
operations.
* Add support for more complex MNNVL topologies with up to 32 NICs per
node.
* If the MNNVL fabric initialization was unsuccessful, NCCL will now fail
by default, so as to avoid inadvertently falling back to a potentially
much slower network transport. Such failures are typically due to a
misconfigured IMEX support on the system. To continue without MNNVL,
restart the job with NCCL_MNNVL_ENABLE=0.
* Fix a potential hang in alltoall-like communication patterns at a scale
of over 80 ranks.
* Make NCCL_P2P_DISABLE=1 imply NCCL_MNNVL_ENABLE=0 (so the latter no
longer needs to be specified on MNNVL systems).
* Fix an initialization failure when NCCL_TOPO_FILE is used on MNNVL
systems.
* Fix the graph search to exclude non-local NICs.
* Fix the SHM transport to use fabric handles on MNNVL systems.
NIC Fusion improvements
* Disable the creation of fused NICs for physical devices that haven't
been merged.
* Flatten multiple ports to a single PCI device within the internal IB
plugin and reparent dual-port NICs under the first PCI parent. If the
parent is not a PCI switch, PCI devices for fused NICs won't be
duplicated.
* Route traffic on GB200-CX8 systems through DirectNIC, not the host
interface.
Improve support for platforms with C2C connectivity (e.g., GB200)
* Enable GPUDirect RDMA for the NICs by default.
* Add support for P2C (PXN over C2C) and the LL128 protocol.
Extend NCCL fault tolerance in multithreaded scenarios
* Support the creation of multiple nonblocking communicators within a
single group and polling in parallel for the completion using multiple
threads (one per communicator).
Enable ncclImplicitOrderLaunch for CUDA 12.9+
* This can potentially speed up NCCL_IMPLICIT_LAUNCH_ORDER.
Improve the netSocket transport latency and control
* Provide finer control over the size of the socket send/receive buffers,
the task size, and the number of sockets that a single peer can open.
* Add support for the inlining of small messages behind the header when
using multiple sockets per connection.
Improve the readability of the CPU affinity in the debug output
* Print it as a range string rather than a bitmask.
Fix a potential race condition in graph execution
* A contention could arise when mixing graph and non-graph execution.
Improve PXN connection code
* Avoid duplicate and unused connections.
RAS fixes
* Fix a memory corruption at job termination time in case of a previously
failed initialization of a RAS socket connection.
* Fix a race condition leading to a crash when generating a RAS report
during communicator initialization (Issues #1669, #1718).
* Fix a potential race condition when gathering data for a RAS status
report.
Fix a potential memory corruption in ncclCommSplit()
* Memory could get corrupted when resource sharing was in use and the size
of the NVLink domain in the new communicator was smaller than in the old
one.
Fix asynchronous graph upload
* Fix a small memory leak.
* Fix oversychronization.
Add a check for out-of-memory conditions in ncclMemAlloc()
Clean up the NCCL socket code
* accept() will retry also if just reading the magic failed (Issue #1613).
* connect() will retry also if poll() did not return a POLLOUT event
(Issue #1618).
* Add error checking in a few instances (Issue #1539).
* Fix the loop condition in ncclFindInterfaceMatchSubnet() (Issue #1574).
* Clean up the debug output, downgrading WARN messages to INFO in
non-critical cases, and printing the peer's address where relevant.
Switch NCCL_DEBUG_FILE to line buffering
* This should help avoid mixed-up partial output lines in multithreaded
cases.
Other minor fixes
* Improve the checks for buffer overflows in the graph code (Issue #1585).
* Extend logging and state clearing to all four events in the internal IB
plugin (Issue #1650).
* Fix the error path in case IB communication is not ready (Issue #1489).
* Add ECE logging for IB fabric.
* Fix various minor issues in the graph module (Issue #1635).
* Clean up the debug output in the graph code, downgrading WARN messages
to INFO in non-critical cases.
* Add a missing argument to a directSend() call (Issue #1628).
* Remove duplicate code in sendProxySetup() (Issue #1420).
* Fix the order of arguments of cudaDeviceCanAccessPeer() (Issue #1507).
* Fix compiler warnings with GCC 14.
* Fix a typo in a comment (Issue #1236).
* Update LL128 elems per thread
* Precompute ix[g] in LL128 prim
* Make Threadthreshold part of tuning models
* Ignore channel tuning when channels are env controlled
* Tune LL128 max limit for AG
* Tune LL128 max limit for RS
* Retune AR LL128 limits due to changes
* Update CHANGELOG.md
---------
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
Minimize the performance impact of the device kernel profiling support when
the profiler plugin is not loaded.
Reduce the overheads of CUDA graph capturing, which increased in NCCL
2.26.2 for large graphs.
Fix the exchange of enhanced connection establishment (ECE) options to
address potential slowdowns on networks utilizing RoCE.
Test if cuMem host allocations work and if not, disable them. Enabled by
default since NCCL 2.24 if the CUDA driver version is at least 12.6, such
allocations rely on NUMA support, which is by default not available under
Docker. We recommend invoking Docker with "--cap-add SYS_NICE" to enable
it.
Fix an initialization error when running with NCCL_NET_GDR_C2C=1 on
multiple MNNVL domains with non-uniform network configurations across
nodes.
Fix the printing of sub-seconds in the debug log when using a custom
NCCL_DEBUG_TIMESTAMP_FORMAT setting.
* Enabling LL128 by default on MI300
* Add missing CUDACHECK
* Adjust BW correction factors to fix the Tree->Ring switching point
* Refactor and add ll128 AR logarithmic factor to tuning models
* Move RCCL tuning changes to a separate file
* Use enum for tunable indexing
* Use explicit indexing in tuning models to avoid mismatch issues
* Place rcclGetSizePerRank in a function
* Remove HIP ifdef for rccl-only call
---------
Co-authored-by: Mustafa Abduljabbar <mustafa.abduljabbar@amd.com>
* removed gfx940 and gfx941
* removed gfx940 and gfx941
* Update "gfx94" to "gfx942" in init.cc
* Updated remaining "gfx94" updates to "gfx942"
* Update filenames and variables from gfx940 to gfx942
---------
Co-authored-by: akolliasAMD <akollias@amd.com>
Profiler improvements
* Add events for CUDA kernel start and end.
* Allow network plugins to generate profiling events
* Enable profiling on a per-operation basis, rather than per-communicator.
* Add support for graph capturing.
Add implicit launch order
* Allow to prevent deadlocks when using multiple NCCL communicators per
device by implicitly ordering NCCL operations using the host program
order. Disabled by default, set NCCL_LAUNCH_ORDER_IMPLICIT=1 to enable.
* Add a complementary mechanism to detect host threads racing to launch
to the same device. Enabled by default, set NCCL_LAUNCH_RACE_FATAL=0 to
disable.
Optimize the PAT algorithm
* Separate the computation and execution of PAT steps on different warps,
allowing to run up to 16 PAT steps in parallel to significantly
accelerate PAT and reduce its linear part.
Add support for setting QoS per communicator
* Add a new trafficClass field to the communicator configuration, to
allow the application to select a particular traffic class for a
given communicator. The meaning of the traffic class is
network-specific and should be set in accordance with the network
configuration.
* For the IB/RoCE plugin, existing config variables such as NCCL_IB_SL
and NCCL_IB_TC take precedence.
Allow to enable GPU Direct RDMA specifically on C2C platforms
* Disabled by default, set NCCL_NET_GDR_C2C=1 to enable.
Do not disable user buffer registration unless PXN is really used
* Only disable UB when a communicator has more than one rank per
node on any node.
RAS subsystem improvements
* Report operation counts separately for each collective operation type.
* Provide details about missing communicator ranks and reliably
distinguish ranks that are no longer a given communicator's members
(now reported as NOCOMM) from those that failed to respond.
Add support for timestamps to NCCL diagnostic messages
* On by default for WARN messages; NCCL_DEBUG_TIMESTAMP_LEVELS can be
used to enable them for other debug levels as well.
* The format can be changed using the NCCL_DEBUG_TIMESTAMP_FORMAT config
variable.
Reduce the memory usage with NVLink SHARP (NVLS)
* Potentially save hundreds of MBs of device memory, considering the
multicast buffer size granularity separately from the address alignment.
Update performance tuning for recent Intel CPUs
* Improve algorithm/protocol selection on recent CPUs such as Emerald
Rapids and Sapphire Rapids.
Improve channel scheduling when mixing LL and Simple operations.
* Make LL operations account for 4x more traffic to ensure LL and simple
operations complete at the same time.
Refactor the plugin code
* Clean up and harmonize the support code across the network, tuner,
and profiler plugins.
Add support for comment lines (starting with #) in the nccl.conf file
* Issue #1540.
Make user buffer registration problems print an INFO instead of a WARN.
Drop support for network plugin interface version 5.
Fix a race condition with split-shared communicators
* NCCL could hang during connection setup if multiple communicators
were grouped together that share resources.
Fix a performance regression when using NCCL_CROSS_NIC=1
* NCCL would unnecessarily alternate rings, breaking the GPU-NIC
associations.
Make GID index detection code more resilient
* Dynamic GID detection code was giving up too soon if the
detected index was not available (e.g., wasn't mapped to the
container's sysfs).
* Issues #1538, #1573.
Fix a race condition with non-blocking operation
* Fix issue when creating a non-blocking communicator after a non-
blocking collective operation on another communicator.
Fix shared memory usage on recent Blackwell GPUs.
* Issues NVIDIA/nccl-tests#287, NVIDIA/nccl-tests#291, #1637.
Fix an error with NIC fusion and IB SHARP when recreating communicators
* Disable the unloading of network plugins
Make the auto-merge failures in the NIC fusion non-fatal
* This could happen when trying to merge IB and RoCE devices.
Fixes to ncclCommAbort
* Fix hangs due to the progress thread spinning indefinitely on the
network progress.
* Reduce the abort time by up to two orders of magnitude.
Fix a crash when libnccl.so was dynamically unloaded
* The RAS subsystem was missing a clean-up handler.
Fix a hang if the network plugin's test() call returns an error.
Fix a hang on heterogeneous architectures
* Ensure we harmonize the tuning to avoid different tuning choices,
causing a hang.
Fix double-free on failed ncclCommInitRank and ncclCommFinalize.
Fix a potential list traversal bug during a group launch of multiple
communicators
* Issue #1599.
Unify the handling of NCCL configuration variables
* Under rare circumstances, some variables specified in the config file
could be ignored.
Avinash
alexander-sannikov
Mustafa Abduljabbar
isaki001
gilbertlee-amd
Rahul Vaidya
BertanDogancay
Wenkai Du
Arm Patinyasakdikul
Nilesh M Negi
Kamil Iskra
Bertan Dogancay
Pedram Alizadeh
corey-derochie-amd