Visible to Intel only — GUID: GUID-0B76C685-E932-4142-8A00-C83F503B47B8
Fusion of Communication Operations
In some cases, it may be beneficial to postpone execution of communication operations and execute them all together as a single operation in a batch mode. This can reduce operation setup overhead and improve interconnect saturation.
oneCCL provides several knobs to enable and control such optimization:
The fusion is enabled by CCL_FUSION.
The advanced configuration is controlled by:
CCL_FUSION_BYTES_THRESHOLD
CCL_FUSION_COUNT_THRESHOLD
CCL_FUSION_CYCLE_MS
NOTE:
For now, this functionality is supported for allreduce operations only.