PReCCL: Performant and Resilient Collective Communication via Integrated Inband Telemetry and Workload Reallocation
PReCCL is a drop-in NCCL replacement that combines software inband telemetry with cross-VT workload reallocation, and implements in-band monitoring within the CCL, and precisely measures the stall counts of each VT, and piggybacks the telemetry meta-data on existing collective traffic.