[gspark-etri]

Background

The README demo from #9 was recorded on an idle cluster: SERVING 0/25, empty Traffic/Serving
views and idle devices (#10). This PR re-records it while a model actually serves traffic.

Changes

• assets/demo.gif, assets/demo.cast: new 44 s recording (120×34, 438 frames). During the
  recording a temporary prefill/decode deployment (1 prefill + 1 decode RNGD card, Llama-3.1-8B,
  shm KV transfer) served behind the llm-d gateway route while rm-bench replayed the Azure
  conversation trace through it at about 1 req/s (--api completions).
• scripts/record-demo.py: new key script. The tour now dwells on views that move under load:
  Overview (7 s) → Traffic: Flow (5 s) → Serving (4 s) → Serving: Perf (8 s, device
  utilisation over time) → Infra: Nodes (4 s) → Infra: Devices sorted busiest first with O
  (9 s, includes the ai-rm AcceleratorProfile / metrics / placement panels) → Overview (7 s).
  Deploy, Setup, Topology and the EPP sub-tab are dropped (nothing changes there under load).
  Still views and sort order only; nothing opens an action menu.
• docs/DEMO.md: steps are now serve + load → record → restore, with an rm-bench example; notes
  that startup readiness-probe failures count as warning events on the Overview status line until
  they expire, and which columns need vLLM/Furiosa-LLM or EPP metrics.
• scripts/record-demo.sh: header comment points at the load step.

Verification

• Recording: LMD_DEMO_ALIASES=<untracked alias file> bash scripts/record-demo.sh on the admin
  host, lmd-top built from 1a3b4f3. Sanitizer: frames: 438, duration: 44.0s, replacements: 3331, distinct IPv4 mapped: 1, leak check passed (exit 0).
• Grep of assets/demo.cast for internal node names, home paths and private IPv4
  (10/8, 172.16/12, 192.168/16): 0 hits. .cast parses as JSON lines (440 lines).
• python tools/check_sensitive.py . (ai-rm copy): clean. Tree-wide grep for internal names: clean.
• GIF 694 KB (< 5 MB). Frames checked as PNG at 3 s / 10 s / 20 s / 33 s / 41 s:
  • Overview: SERVING 2/25, status line 2 model(s) serving, 1 accelerator(s) hot (>80%) (no
    warning from the demo workload; its startup probe events had expired before recording),
    RNGD on npu-node-1 at 24 % with trend.
  • Traffic: Flow: /rngd/llama8… route → llama8b-pd-pool (1/1 endpoints) → gateway pod Running.
  • Serving: Perf: RNGD npu0 utilisation sparkline rising to 73 %.
  • Infra: Devices: RNGD npu0 73 %, 59 °C, 114 W; ai-rm placement panel for llama8b-pd-pool.
• Load (rm-bench, same workload, 25 min run just before the recording): 1744 requests,
  0 failed, ~1.16 req/s, TTFT p90 1066 ms, TPOT p90 288 ms. The recording run used
  --speedup 0.75 (~1 req/s).
• Cluster restored afterwards: load stopped, the three demo deployments back to 0 replicas, no
  pods left, Helm release unchanged (revision 4), airm-placement ConfigMap unchanged (same
  resourceVersion and content hash before and after), 0 RNGD/ATOM allocated on the node.

Compatibility and limits

• The workers in this deployment export only airm_* series, not vLLM/Furiosa-LLM metrics, and
  this pool's endpoint picker is not scraped by Prometheus. So the header req/s, TTFT, E2E
  and the Serving: Perf throughput row still show –, and EPP endpoint decisions are empty.
  Utilisation, power, temperature, serving status and the ai-rm panels move. Showing request
  rate/TTFT needs a vLLM or Furiosa-LLM server with a ServiceMonitor (or the PD worker exporting
  vLLM-compatible names); left for a later re-record. docs/DEMO.md explains this.
• The header shows ⚠8 alert (A): pre-existing cluster alerts (image pulls on other nodes), not
  from the demo workload. Another team's job had an A100 at 100 % at recording time; it shows as is.
• No lmd-top code changes.

Related Issue

Closes #10

🤖 Generated with Claude Code