SimCCL

SimCCL is the collective communication translation layer for the SimAI network simulator. It converts NCCL collective operation decisions (algorithm/protocol/channel/chunk) into FlowModels — structured point-to-point flow descriptions — that drive ns3 packet-level simulation or can be exported as CSV for offline analysis.

SimCCL does NOT implement GPU kernels or network I/O. It translates "what NCCL would decide" into "what flows would result".

Supported Operations

Collective	Algorithms	Notes
AllReduce	Ring, NVLS	NVLS for single-node >=8 GPU
AllGather	Ring, PAT	PAT for multi-node 1-GPU/node
ReduceScatter	Ring, PAT	PAT for multi-node 1-GPU/node
AlltoAll	Ring	—
Broadcast	Ring	—

Note: Tree algorithm constants exist but auto-selection is not triggered on tested hardware (H20). CollNetDirect/CollNetChain are not implemented (require SHARP hardware).

Quick Start (Standalone Mode)

Build

cd SimCCL/src
bash build.sh v2.30 # default v2.30, or pass v2.20

Output: build/simccl-standalone (~300KB, no ns3 dependency, CPU-only)

Note: the standalone build includes astra-sim/system/Common.hh and its CMake references ../../astra-sim-alibabacloud, so SimCCL must sit inside a SimAI checkout (SimAI/SimCCL/ next to SimAI/astra-sim-alibabacloud/). See Installation for the required layout.

Run

# Single collective operation
./build/simccl-standalone --op AllReduce --size 4194304 \
--nRanks 8 --nNodes 1 --gpus_per_node 8 --gpu_type H20
# Workload file mode
./build/simccl-standalone -w ../../example/microAllReduce.txt \
--nRanks 16 --nNodes 2 --gpus_per_node 8 --gpu_type H20

Output: ncclFlowModel_detailed_flows.csv

Full Test Suite

cd SimCCL
bash scripts/run_all.sh

Integration with SimAI

When compiled as part of the full SimAI simulator, SimCCL provides FlowModels to the ns3 network simulation backend:

cd SimAI/
./scripts/build.sh -c ns3

See Integration Guide for details.

Documentation

See docs/README.md for full documentation index.

License

MIT — See LICENSE