bahnland/dodoco-moe-traffic-matrices

DODOCO: MoE AlltoAll Traffic Matrices on Real Text

Per-dispatch AlltoAll traffic matrices (send counts) from five released Mixture-of-Experts checkpoints on 16 GPUs, with the router held at its released weights and real text as input. Each matrix records how many tokens each of the 16 ranks sends to each other rank in one MoE dispatch. The data is meant as workload input for interconnect, collective and topology studies, so that measured MoE traffic can be used without setting up the training stack.

Models

Folder	Checkpoint	Sequence mixer	Experts	Top-k	MoE layers	Hidden size
DeepSeek-MoE-16B	deepseek-ai/deepseek-moe-16b-base	MHA	64	6	27	2048
DeepSeek-V2-Lite	deepseek-ai/DeepSeek-V2-Lite	MLA	64	6	26	2048
Qwen3-30B-A3B	Qwen/Qwen3-30B-A3B	GQA	128	8	48	2048
Nemotron-3-Nano-30B-A3B	nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16	hybrid: Mamba-2 + attention	128	6	23	2688
Qwen3.5-35B-A3B	Qwen/Qwen3.5-35B-A3B-Base	hybrid: Gated DeltaNet + attention	256	8	40	2048

Inputs

Folder	Text
english	English Wikipedia, WikiText-2
opus	English reasoning traces generated by Claude Opus 4.6, assistant turns of TeichAI/Claude-Opus-4.6-Reasoning-887x
romansh	Rumantsch Grischun, monolingual subset of swiss-ai/apertus-pretrain-romansh

Each text is tokenized with the model's own tokenizer. Each cell has one run (rep1).

Setup

• 16 NVIDIA H100 GPUs on 4 nodes, 4 GPUs per node; rank r runs on node r // 4.
• NVIDIA NeMo 26.02 and Megatron-Core 0.16.1, AlltoAll token dispatcher, expert parallelism 16, tensor and pipeline parallelism 1. Experts are placed contiguously: rank r holds experts r * E/16 to (r+1) * E/16 - 1.
• Global batch 16 sequences of 4096 tokens, micro-batch 1: each rank routes 4096 tokens per step.
• Each run takes 15 steps (5 warmup + 10) through the training loop with the router held at its released weights: learning rate and minimum learning rate 0, and for Nemotron also an expert-bias update rate of 0. router_state.csv holds a checksum of every router weight and expert-bias tensor at every step; all are bit-identical within each run, so every step routes with the released router. At the training recipe's own learning rate the router rebalances within a few steps, which a released-checkpoint profile has to avoid.
• The counts come from a forward hook on the dispatcher that reads its input splits, the per-destination token counts the dispatcher already computes.

Files

manifest.csv                 one row per run: model, checkpoint, top-k, input, repetition, layer
                             and expert counts, hidden size, steps, dispatches per layer per step
data/<model>/<input>/rep<n>/
  send_matrices.csv          one row per (dispatch, source rank)
  expert_counts.csv          tokens per expert per layer and step, summed over all 16 ranks
  router_state.csv           router checksums per step (sum, l2, absmax), proof of the freeze
  config.json                run configuration
example.py                   loader and summary statistics, standard library only
SHA256SUMS                   checksums of every data file

send_matrices.csv columns:

Column	Meaning
call	dispatch index within the run, in execution order
step	training step, 1 to 15 (1 to 5 are warmup steps)
pass	forward, or recompute (see below)
layer	MoE layer index, 0-based, counting MoE layers only
src_rank	sending rank
to_rank_0 ... to_rank_15	tokens sent from src_rank to that rank

The 16 rows sharing a call form one 16 x 16 matrix. Combine, the second AlltoAll of each MoE layer, returns the same volumes in the reverse direction (the transposed matrix).

Reading the counts

• Counts are token-expert assignments. A token routed to several experts on the same destination rank is counted once per expert, which is what the dispatcher sends. Each row sums to 4096 x top-k.
• Bytes. Multiply a count by hidden_size x bytes per element of the dispatch precision (2 for bf16) to get the payload.
• Self traffic. The diagonal (src_rank == to_rank_j) is traffic that stays on the GPU.
• Recompute pass. These runs use full activation recomputation, which repeats each layer's forward pass during the backward pass, so every MoE layer is dispatched twice per step. The recompute dispatch is identical to its forward dispatch and is real traffic on the wire. Drop it to model training without recomputation, or inference. The Nemotron runs record one dispatch per MoE layer per step, all forward.
• Other expert-parallel widths. Summing contiguous blocks of ranks gives the full traffic of the same tokens at EP 8, 4 or 2 (coarsen in example.py). For other placements, including more ranks, expert_counts.csv gives each rank's receive total, but not the source-by-destination matrix: the matrices record destination ranks, not destination experts.
• expert_counts.csv counts, per step and MoE layer, the token-expert assignments each expert received from all 16 ranks. Columns: step, layer, expert_0 ... expert_{E-1}. It covers steps 8 to 15 (the tracker sums steps 1 to 7 into one record, which is left out). Counts include the recompute pass, so for the four models with recomputation each row is twice the forward load; the counts of each rank's expert block match the forward receive totals of send_matrices.csv at that step, times the number of dispatches per layer per step.

Not included

Runs on synthetic inputs (uniform-random, Zipf-distributed, shuffled or vocabulary-remapped tokens), runs with the router in training, and per-run timing, energy and network counters.

Citation

@inproceedings{ma2026dodoco,
  author    = {Ma, Bole and Eitzinger, Jan and K{\"o}stler, Harald and Wellein, Gerhard},
  title     = {Diagnosing Overhead in Dispatch Operations: Cross-architecture Observatory},
  booktitle = {Proc. 2026 IEEE Symp. High-Performance Interconnects (HOTI)},
  year      = {2026},
  note      = {arXiv:2605.20982}
}