Environment Variables

Collective algorithms selection

CCL_<coll_name>

Syntax

To set a specific algorithm for the whole message size range:

CCL_<coll_name>=<algo_name>

To set a specific algorithm for a specific message size range:

CCL_<coll_name>="<algo_name_1>[:<size_range_1>][;<algo_name_2>:<size_range_2>][;...]"

Where:

  • <coll_name> is selected from a list of available collective operations (Available collectives).

  • <algo_name> is selected from a list of available algorithms for a specific collective operation (Available algorithms).

  • <size_range> is described by the left and the right size borders in a format <left>-<right>. Size is specified in bytes. Use reserved word max to specify the maximum message size.

oneCCL internally fills algorithm selection table with sensible defaults. User input complements the selection table. To see the actual table values set CCL_LOG_LEVEL=info.

Example

CCL_ALLREDUCE="recursive_doubling:0-8192;rabenseifner:8193-1048576;ring:1048577-max"

Available collectives

Available collective operations (<coll_name>):

  • ALLGATHERV

  • ALLREDUCE

  • ALLTOALL

  • ALLTOALLV

  • BARRIER

  • BCAST

  • REDUCE

  • REDUCE_SCATTER

  • SPARSE_ALLREDUCE

Available algorithms

Available algorithms for each collective operation (<algo_name>):

ALLGATHERV algorithms

direct

Based on MPI_Iallgatherv

naive

Send to all, receive from all

flat

Alltoall-based algorithm

multi_bcast

Series of broadcast operations with different root ranks

ALLREDUCE algorithms

direct

Based on MPI_Iallreduce

rabenseifner

Rabenseifner’s algorithm

starlike

May be beneficial for imbalanced workloads

ring

reduce_scatter + allgather ring. Use CCL_RS_CHUNK_COUNT and CCL_RS_MIN_CHUNK_SIZE to control pipelining on reduce_scatter phase.

ring_rma

reduce_scatter+allgather ring using RMA communications

double_tree

Double-tree algorithm

recursive_doubling

Recursive doubling algorithm

2d

Two-dimensional algorithm (reduce_scatter + allreduce + allgather)

ALLTOALL algorithms

direct

Based on MPI_Ialltoall

naive

Send to all, receive from all

ALLTOALLV algorithms

direct

Based on MPI_Ialltoallv

naive

Send to all, receive from all

BARRIER algorithms

direct

Based on MPI_Ibarrier

ring

Ring-based algorithm

BCAST algorithms

direct

Based on MPI_Ibcast

ring

Ring

double_tree

Double-tree algorithm

naive

Send to all from root rank

REDUCE algorithms

direct

Based on MPI_Ireduce

rabenseifner

Rabenseifner’s algorithm

tree

Tree algorithm

double_tree

Double-tree algorithm

REDUCE_SCATTER algorithms

direct

Based on MPI_Ireduce_scatter_block

ring

Use CCL_RS_CHUNK_COUNT and CCL_RS_MIN_CHUNK_SIZE to control pipelining.

SPARSE_ALLREDUCE algorithms

ring

Ring-allreduce based algorithm

mask

Mask matrix based algorithm

allgatherv

3-allgatherv based algorithm

Note

WARNING: ccl::sparse_allreduce is experimental and subject to change.

CCL_RS_CHUNK_COUNT

Syntax

CCL_RS_CHUNK_COUNT=<value>

Arguments

<value>

Description

COUNT

Maximum number of chunks.

Description

Set this environment variable to specify maximum number of chunks for reduce_scatter phase in ring allreduce.

CCL_RS_MIN_CHUNK_SIZE

Syntax

CCL_RS_MIN_CHUNK_SIZE=<value>

Arguments

<value>

Description

SIZE

Minimum number of bytes in chunk.

Description

Set this environment variable to specify minimum number of bytes in chunk for reduce_scatter phase in ring allreduce. Affects actual value of CCL_RS_CHUNK_COUNT.

Fusion

CCL_FUSION

Syntax

CCL_FUSION=<value>

Arguments

<value>

Description

1

Enable fusion of collective operations

0

Disable fusion of collective operations (default)

Description

Set this environment variable to control fusion of collective operations. The real fusion depends on additional settings described below.

CCL_FUSION_BYTES_THRESHOLD

Syntax

CCL_FUSION_BYTES_THRESHOLD=<value>

Arguments

<value>

Description

SIZE

Bytes threshold for a collective operation. If the size of a communication buffer in bytes is less than or equal to SIZE, then oneCCL fuses this operation with the other ones.

Description

Set this environment variable to specify the threshold of the number of bytes for a collective operation to be fused.

CCL_FUSION_COUNT_THRESHOLD

Syntax

CCL_FUSION_COUNT_THRESHOLD=<value>

Arguments

<value>

Description

COUNT

The threshold for the number of collective operations. oneCCL can fuse together no more than COUNT operations at a time.

Description

Set this environment variable to specify count threshold for a collective operation to be fused.

CCL_FUSION_CYCLE_MS

Syntax

CCL_FUSION_CYCLE_MS=<value>

Arguments

<value>

Description

MS

The frequency of checking for collectives operations to be fused, in milliseconds:

  • Small MS value can improve latency.

  • Large MS value can help to fuse larger number of operations at a time.

Description

Set this environment variable to specify the frequency of checking for collectives operations to be fused.

ATL

CCL_ATL_TRANSPORT

Syntax

CCL_ATL_TRANSPORT=<value>

Arguments

<value>

Description

mpi

MPI transport (default).

ofi

OFI (Libfabric*) transport.

Description

Set this environment variable to select the transport for inter-process communications.

CCL_ATL_HMEM

Syntax

CCL_ATL_HMEM=<value>

Arguments

<value>

Description

1

Enable heterogeneous memory support on the transport layer.

0

Disable heterogeneous memory support on the transport layer (default).

Description

Set this environment variable to enable handling of HMEM/GPU buffers by the transport layer. The actual HMEM support depends on the limitations on the transport level and system configuration.

CCL_UNORDERED_COLL

Syntax

CCL_UNORDERED_COLL=<value>

Arguments

<value>

Description

1

Enable execution of unordered collectives. You have to additionally specify match_id.

0

Disable execution of unordered collectives (default).

Description

Set this environment variable to enable execution of unordered collective operations on different nodes.

CCL_PRIORITY

Syntax

CCL_PRIORITY=<value>

Arguments

<value>

Description

direct

You have to explicitly specify priority using priority.

lifo

Priority is implicitly increased on each collective call. You do not have to specify priority.

none

Disable prioritization (default).

Description

Set this environment variable to control priority mode of collective operations.

CCL_WORKER_COUNT

Syntax

CCL_WORKER_COUNT=<value>

Arguments

<value>

Description

N

The number of worker threads for oneCCL rank (1 if not specified).

Description

Set this environment variable to specify the number of oneCCL worker threads.

CCL_WORKER_AFFINITY

Syntax

CCL_WORKER_AFFINITY=<cpulist>

Arguments

<cpulist>

Description

auto

Workers are automatically pinned to last cores of pin domain. Pin domain depends from process launcher. If mpirun from oneCCL package is used then pin domain is MPI process pin domain. Otherwise, pin domain is all cores on the node.

<cpulist>

A comma-separated list of core numbers and/or ranges of core numbers for all local workers, one number per worker. The i-th local worker is pinned to the i-th core in the list. For example <a>,<b>-<c> defines list of cores contaning core with number <a> and range of cores with numbers from <b> to <c>. The number should not exceed the number of cores available on the system.

Description

Set this environment variable to specify cpu affinity for oneCCL worker threads.

CCL_WORKER_MEM_AFFINITY

Syntax

CCL_WORKER_MEM_AFFINITY=<nodelist>

Arguments

<nodelist>

Description

auto

Workers are automatically pinned to NUMA nodes that correspond to CPU affinity of workers.

<nodelist>

A comma-separated list of NUMA node numbers for all local workers, one number per worker. The i-th local worker is pinned to the i-th NUMA node in the list. The number should not exceed the number of NUMA nodes available on the system.

Description

Set this environment variable to specify memory affinity for oneCCL worker threads.

CCL_LOG_LEVEL

Syntax

CCL_LOG_LEVEL=<value>

Arguments

<value>

error

warn (default)

info

debug

trace

Description

Set this environment variable to control logging level.

CCL_MAX_SHORT_SIZE

Syntax

CCL_MAX_SHORT_SIZE=<value>

Arguments

<value>

Description

SIZE

Bytes threshold for a collective operation (0 if not specified). If the size of a communication buffer in bytes is less than or equal to SIZE, then oneCCL does not split operation between workers. Applicable for allreduce, reduce and broadcast.

Description

Set this environment variable to specify the threshold of the number of bytes for a collective operation to be split.

Multi-NIC

CCL_MNIC, CCL_MNIC_NAME and CCL_MNIC_COUNT define filters to select multiple NICs. oneCCL workers will be pinned on selected NICs in a round-robin way.

CCL_MNIC

Syntax

CCL_MNIC=<value>

Arguments

<value>

Description

global

Select all NICs available on the node.

local

Select all NICs local for the NUMA node that corresponds to process pinning.

none

Disable special NIC selection, use a single default NIC (default).

Description

Set this environment variable to control multi-NIC selection by NIC locality.

CCL_MNIC_NAME

Syntax

CCL_MNIC_NAME=<namelist>

Arguments

<namelist>

Description

<namelist>

A comma-separated list of NIC full names or prefixes to filter NICs. Use the ^ symbol to exclude NICs starting with the specified prefixes. For example, if you provide a list mlx5_0,mlx5_1,^mlx5_2, NICs with the names mlx5_0 and mlx5_1 will be selected, while mlx5_2 will be excluded from the selection.

Description

Set this environment variable to control multi-NIC selection by NIC names.

CCL_MNIC_COUNT

Syntax

CCL_MNIC_COUNT=<value>

Arguments

<value>

Description

N

The maximum number of NICs that should be selected for oneCCL workers. If not specified then equal to the number of oneCCL workers.

Description

Set this environment variable to specify the maximum number of NICs to be selected. The actual number of NICs selected may be smaller due to limitations on transport level or system configuration.