weili-0234/mpnd-routing-m27-2p4d-edgebench

MpNd routing campaign — 2P4D MiniMax-M2.7, decode-policy ablation (EdgeBench strict replay)

Complete raw traces + analysis artifacts for the mpnd-routing-m27 campaign (research-common H100, 2026-08-11/12). Six steady-state cells — decode worker selection policy {power_of_two, cache_aware} × fleet-total concurrency {32, 48, 64} — plus a two-cell 192-session overload annex, on a 2-prefill / 4-decode SGLang PD-disaggregation fleet (tp8+ep8 per worker, 48×H100, MiniMax-M2.7 FP8-blockwise, BF16 KV, page 64, HiCache both roles ratio 3.0 for the t-cells / 2.0 for the annex, decode radix cache + delta KV transfer, mooncake RDMA). Router: sgl-model-gateway, sticky-session prefill (manual policy + X-SMG-Routing-Key, min_group); ONLY --decode-policy differs between arms. Engine fork commit fe02c2ca39 (sglang-private, branch weili/mpnd-m27-routing). Workload: trie strict replay of the bit-exact EdgeBench M2.7 corpus (137 traces, forced output tokens, per-session-instance cache salt, seed 20260810), 3600 s admission + ≤1800 s drain per cell. Bit-exact audit violations: 0 in every cell.

Corpus (needed to REPLAY, not to recompute metrics): https://huggingface.co/datasets/weili-0234/edgebench-bit-exact-m27

Layout

cells/<cell_id>/                    # per-cell raw bundle (the settlement input)
  lifecycle/client.jsonl            #   per-request CLIENT_ADMIT/FIRST_TOKEN/COMPLETE/
                                    #   REQUEST_FAILED events, ns timestamps, token counts,
                                    #   real+ideal cache-hit fields, session/trace hashes
  timed-result.json                 #   trie whole-run result incl. per-request samples
  metrics/<w>.loads.openmetrics     #   per-worker /v1/loads series, 5 s cadence (imbalance truth)
  metrics/<w>.openmetrics           #   full per-worker Prometheus dumps
  preflight/, transition-receipt.json, cell-config.json, timing.json,
  phase-markers.jsonl, process-epochs.json, corpus-receipt.json
cells-archive/                      # aborted/degraded cell forensics (incidents)
workers/<w>/server.log              # engine logs: DECODE_PREALLOC per-request anchors,
                                    #   Prefill batch lines, KV pool sizes (1 s timestamps)
workers/<w>/requests/               # per-request engine record dumps
workers/<w>/gpu-util.csv, host-memory-kib.csv
workers-archive-9-pre-3p0-respec/   # annex-era (hicache 2.0) worker logs incl.
                                    #   d0-gen9-hang/ (pre-incident d0 log)
router/logs/gen-<N>.log             # sgl-model-gateway logs: PD_DISPATCH per-request
                                    #   prefill/decode pair attribution (+ retries)
control*/, controller/, chain-markers.jsonl, corpus-receipt.json
analysis/                           # everything below computed FROM the raw layer
  settle_mpnd_cell.py               #   canonical settlement (window [T0+1800,3600))
  mirror_cell.sh                    #   cluster->local mirror helper
  settled/<cell>/settlement.{json,txt} + imbalance CSVs + per-decode CSVs
  report/mpnd-m27-report.html       #   self-contained interactive report
  report/extract_report_data.py     #   per-request TTFT/E2E/TPS extraction
  report/extract_imbalance_data.py  #   Jain/max-share imbalance time series
  report/prealloc-all.csv, prealloc-d0-archive.csv, prefill-thr.csv
  report/ttft-breakdown-all.txt     #   TTFT segment percentiles (NB: contains
                                    #   throughput-model ESTIMATES for compute/transfer)
  report/ttft-breakdown-pct.json    #   measured 2-segment percentile breakdown
  incidents/d0-hang-*.txt           #   py-spy stacks of the HiCache write-side deadlock

Recomputing metrics

Settlement (any cell): python3 analysis/settle_mpnd_cell.py --cell-dir cells/<cell_id> --out-dir out/ — produces TTFT/E2E/TPS/TPOT distributions, real-vs-ideal cache hit, pool imbalance (loads series), per-decode-worker attribution (joins router/logs/gen-N.log via the cell's transition-receipt router_generation). Report data: analysis/report/extract_report_data.py (percentile curves, user-defined TPS: gen=ct/e2e, decode=ct/(e2e−ttft), sys=Σct/window) and extract_imbalance_data.py (30 s-bucket Jain fairness / max-share). TTFT two-segment breakdown: join lifecycle ns timestamps to DECODE_PREALLOC lines in workers/*/server.log per rid occurrence within [admit−2 s, first_token+2 s] (rids repeat across cells — never join globally); ±1 s log quantization; cross-role joins by bootstrap_room, not rid.

Headline results (steady window, floor-rank percentiles)

cell	TTFT p50/p99 (s)	wTPS	TPOT p50	real-hit (eff)	new tok (M)
p2-t32	1.65 / 22.6*	99.9	11.5	97.1% (98.9%)	4.03
ca-t32	1.52 / 36.8*	100.9	11.0	98.1% (99.99%)	2.58
p2-t48	2.58 / 26.4*	77.5	14.7	96.7% (98.6%)	5.53
ca-t48	2.87 / 90.9*	79.3	13.9	97.8% (99.9%)	3.61
p2-t64	5.74 / 48.3*	68.4	16.4	96.7% (98.6%)	6.14
ca-t64	6.83 / 58.9*	73.8	14.9	98.0% (99.9%)	3.67

(*p99 from the percentile-bin means of the breakdown data.) Verdict: cache_aware always wins cache economy (−35-40% new prompt tokens); latency parity at t32/t48; at t64 power_of_two wins TTFT (tail −41%) while cache_aware wins decode TPS/TPOT — the concentration penalty is localized to the decode KV-prealloc admission segment. Overload annex (c32 cells, 192 sessions, hicache 2.0) flips: cache_aware TTFT −37%, weighted TPS −21%.

Known engine issue captured here: HiCache write-side collective deadlock (flush_write_through_acks → _all_reduce_attn_groups) hung decode d0 at t64; py-spy evidence in analysis/incidents/; the degraded cell is in cells-archive/mpnd-power_of_two-t64-d0hang-degraded/.