* 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>
* Support pipelining codegen and template specialization
* Support ReduceCopy pipelining for AllReduce, ReduceScatter, and Reduce (currently enabled for bfloat16)
* Remove need for FUNC_INDEX_TOTAL
* Add pipeline field to device function key construction logic
* Avoid unneeded codegen for LL/LL64 kernels
* Modify conditions and add pipeline dtypes env
* Optimize selection for both gfx942 and gfx950
* Increase pipeline bitfield width
* Use __forceinline__ for all device functions
* Realign reduceCopy with original form
* Add opt-out option to enable perf debugs
* Remove force-reduce-pipelining option from README
* Update CHANGELOG.md
---------
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
Leverages the traits of extended-scope fine-grain memory to get rid of a device-scope acquire-release fence. This improves throughput for single node workloads on gfx942 and gfx950 for some input sizes (e.g., ~32 MiB to about 256 MiB) when using the simple protocol. Multinode workloads on MI300X see a smaller but statistically significant uplift for some message sizes. Runtime disablement is supported via setting the environment variable RCCL_GFX942_CHEAP_FENCE_ON to 0.
Prevent initialization failures in certain configurations when attempting
to load fp8-specific symmetric multicast kernels on GPUs older than
Blackwell.
* Support fused all reduce and elementwise operations
Add additional "acc" parameter to RCCL Replayer logs
Add flag which indicates availability of new API
* Fix Recorder json parsing
* Remove unreachable code
* Remove extra acc pointer check
* .
* Revert "[DEVICE] Adding ability to choose unroll factor at runtime (#1734)"
This reverts commit 9d72be7b2f.
* Use noinline to reduce kernels linking time
* Don't use noinline for gfx942 and gfx950 to avoid perf regression
---------
Co-authored-by: AtlantaPepsi <timhu102@amd.com>
Co-authored-by: BertanDogancay <bertan.dogancay@gmail.com>
Boosts single node bfloat16 allreduce performance by up to 20% for some data sizes and provides gating with the RCCL_GFX942_CHEAP_FENCE_OFF environment variable
This feature tracks the proxy events and status of each send/recv op. ProxyTrace keeps a fixed number of active ops in host mem and dumps the status of each op when the program crashes or hangs.
* 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>
* Fix when more than 64 channels are used for multi-collective group calls
* Update CHANGELOG.md
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
---------
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
* Internal RCCL/NCCL functionality exposed when RCCL_EXPOSE_STATIC is enabled
* Algo/protocol/max channels can be obtained with the new RCCL API
* Introduce rccl_static and rccl_static_inline macros to work around invisible functions in core source files like enqueue.cc
* Add usage example in topo-explorer tool
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>
* [SRC] Enable unroll=1 for gfx950
* Fix typo from rebase in generate.py
* Support for unroll=1 and gfx90a when building for all GPU targets
---------
Signed-off-by: nileshnegi <Nilesh.Negi@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.
* Fix collective trace
* Use nontemporal for st_global
* Fix previous commit
* Add HDP flush to data receive path
* Fix previous commit
* Control flushing by NCCL_NET_FORCE_FLUSH and RCCL_NET_HDP_FLUSH
* Introduce RCCL_NET_HDP_FLUSH and RCCL_NET_GDR_FLUSH
Both are on by default. Turn both off will skip all flush will likely
result in data error.
* Enable GDR copy by default
* Remove GDR flush env var because it is disabled by GDC flush
* Output kernel collective trace at comm destroy by default
* Limit kernel timeout messages to 100
* Use system relaxed atomic for loadInt
* Refine timeout messages and use atomic for setting offset from CPU
* Add kernel trace for barrier timeout
* Add backup barrier to avoid race in atomicAdd
* Use different counters for different warps
* Rework barrier implementation
* Fix for other GFX
* Use __hip_atomic_store and __hip_atomic_load
* Fix bug in previous commit
* Don't reset barrier values in running kernel
* Update trace format
* Fix typo
* Switch back to hip_atomic_fetch_add
* Use same barrier implementation for all GFX
* Remove extra threadfence
* Turn off HDP flush by default
Please use RCCL_NET_HDP_FLUSH=1 to switch on HDP flush
* Remove unnecessary changes from alterative barrier implementation
* Added back __threadfence_block
* Revert back to threadfence for gfx other than gfx94x
Add Blackwell/SM100 support
* Add compilation for sm100
* Add graph search speeds for Blackwell
* Optimize graph search to converge on large NVLink domains
* Limit NVLS heads to 32
* Increase various limits to fit large NVLink domains
* Add extra checks for IMEX setup, needed for MNNVL
* Increase MAXCHANNELS to 64
Extend NVTX instrumentation to track NCCL communicators
* Add communicator ID to NVTX traces to allow for correlation
between ranks.
RAS fixes
isaki001
mberenjk
BertanDogancay
Mustafa Abduljabbar
Bertan Dogancay
alex-breslow-amd
Avinash
ycui1984
Kamil Iskra
Wenkai Du
Nilesh M Negi
Arm Patinyasakdikul
Dingming Wu
Pedram Alizadeh
gilbertlee-amd
Tim
Sylvain Jeaugey