PhD Edge Telemetry Lab & Paper Pipeline MCP

Magnus Cruz & Matheus Torquato — CISUC / DEI, University of Coimbra • Sync: Active
Rule 5: SYNCHRONIZED (11 Campaigns) JSON Status Paper Pipeline MCP → Literature Survey MCP →

1. Edge Node Telemetry Readiness (ubuntu@ubuntu-llm (192.168.0.205)) — Agenda A4 Verification

Hardware: 13th Gen Intel(R) Core(TM) i5-13500 (20 Threads: 6P + 8E, 4.80 GHz max, governor: performance)
OUT-OF-BAND WALL POWER (1 Hz)
Tapo P115 @ 192.168.0.200 (1 Hz out-of-band telemetry — ACTIVE: 34 W idle / 155 W load)
IN-BAND CPU PACKAGE TEMPERATURE (1 Hz)
lm-sensors coretemp Package id 0 (1 Hz — ACTIVE: 25.0°C idle / 66.0°C–69.0°C load)
AMBIENT TEMPERATURE & HUMIDITY (BLE)
ambient-logger-ble.service -> /home/ubuntu/ambient_ble.csv (ACTIVE: 22.80°C–23.00°C, 35.0%–48.0% RH)

2. Active Longitudinal Campaign: Experiment D: 48h Extended Mechanism Isolation & Multi-Horizon Rejuvenation

llama3.2:3b-instruct-q4_K_M (8 threads, capped 3500 MHz)
Inferences Done 21,840 Burn-in: 16,137 inf (24h)
Current VmRSS 3703.1 MB Initial: 3658.3 MB (+44.8 MB)
Heap Dirty Fraction 98.5% Dirty: 3675.5 MB (smaps_rollup)
Decode TPS 24.46 Aging Decay: -1.22%
Phase & Layer Isolation Protocol:
• Status: Phase 3: Post-Intervention (L2 Page Cache Drop - POSIX_FADV_DONTNEED / drop_caches=3)
• L1 Thermal Relaxation: 60s idle paused inference → 0.5% recovery (-0.24 MB). Thermal decay ruled out.
• L2 Page Cache Drop: POSIX_FADV_DONTNEED / drop_caches → 0% recovery. Proves memory accumulation is 100% heap dirty.
• L3 Runtime Warm Restart: Complete state rejuvenation with energy cost ΔErejuv = 84.7 J, parameterizing Paper 4's closed form.

3. Experiment 1: Multi-Model Rejuvenation Energy Cost Calibration (80 Cycles = 100%)

Model Intervention Action Action (s) Reload + 1st Infer (s) Total Cycle (s) Action Energy (J) Total Cycle Energy (J) Post-Action Decode Process Verification
qwen2.5:0.5b runner_restart_warm 0.927 s 5.316 s 6.243 s 33.6 J 521.6 J 107.30 tps PID Rotated (5/5)
qwen2.5:0.5b page_cache_drop 0.121 s 3.712 s 3.833 s 10.3 J 366.0 J 107.28 tps Same PID (0/5)
qwen2.5:0.5b runner_restart_cold 0.855 s 6.768 s 7.623 s 45.8 J 580.3 J 106.87 tps PID Rotated (5/5)
qwen2.5:0.5b idle_pause_control 6.885 s 4.588 s 11.473 s 244.8 J 688.4 J 106.02 tps Same PID (0/5)
phi3:mini runner_restart_warm 6.324 s 22.425 s 28.749 s 759.1 J 4007.5 J 23.26 tps PID Rotated (5/5)
phi3:mini page_cache_drop 0.048 s 67.966 s 68.014 s 4.1 J 9181.9 J 21.10 tps Same PID (0/5)
phi3:mini runner_restart_cold 6.888 s 40.052 s 46.940 s 585.5 J 6336.9 J 22.38 tps PID Rotated (5/5)
phi3:mini idle_pause_control 27.107 s 23.218 s 50.326 s 2304.1 J 6794.0 J 22.76 tps Same PID (0/5)
llama3.1:8b runner_restart_warm 3.470 s 47.493 s 50.962 s 294.9 J 6879.9 J 12.26 tps PID Rotated (5/5)
llama3.1:8b page_cache_drop 0.064 s 48.373 s 48.437 s 5.5 J 6539.0 J 12.21 tps Same PID (0/5)
llama3.1:8b runner_restart_cold 4.156 s 74.429 s 78.585 s 353.3 J 10609.0 J 12.16 tps PID Rotated (5/5)
llama3.1:8b idle_pause_control 50.193 s 46.144 s 96.337 s 4266.4 J 13005.4 J 12.16 tps Same PID (0/5)
deepseek-v2:lite runner_restart_warm 2.575 s 20.640 s 23.215 s 218.9 J 3134.1 J 29.02 tps PID Rotated (5/5)
deepseek-v2:lite page_cache_drop 0.171 s 16.009 s 16.181 s 14.6 J 2184.4 J 29.06 tps Same PID (0/5)
deepseek-v2:lite runner_restart_cold 2.342 s 60.800 s 63.141 s 199.0 J 8524.1 J 29.05 tps PID Rotated (5/5)
deepseek-v2:lite idle_pause_control 20.159 s 14.361 s 34.520 s 1713.5 J 4660.2 J 29.44 tps Same PID (0/5)

4. Four-Paper PhD Track Synchronization & Pipeline Status

PAPER1-IOTJ \TBD: 0
Slowing Down to Save: A Longitudinal, Carbon-Aware Evaluation of Large Language Model Inference on CPU-Only Edge Nodes
IEEE Internet of Things Journal (IoT-J)
SUBMITTED & UNDER PEER REVIEW (12 Pages Exact)
Cites PAPER1: 0 | Cites PAPER3: 0
PAPER2-AGING-POSITION \TBD: 0
Software Aging and Rejuvenation for Large Language Model Inference on IoT and Edge Nodes: A Position and Research Agenda
IEEE Trans. on Emerging Topics in Computing (TETC) SI WoSAR 2026 (Q1, IF 4.8)
READY FOR REVIEW (5 Pages Exact, 12/12 Review Notes Resolved)
Cites PAPER1: 9 | Cites PAPER3: 0
PAPER3-MECHANISM \TBD: 0
Layer Localization and Mechanism Isolation of Performance Decay in Sustained Edge LLM Inference
IEEE/IFIP DSN 2027 Disrupt Track (Berlin)
DRAFT COMPLETE — IEEE ABSTRACT & MULTI-CAMPAIGN RESULTS (7 Pages Exact)
Cites PAPER1: 7 | Cites PAPER3: 0
PAPER4-FRAMEWORK \TBD: 0
Energy- and Carbon-Aware Software Rejuvenation Scheduling for Edge LLM Services
IEEE Trans. Green Communications and Networking (TGCN) / IEEE TSE
ARCHITECTURE FORMULATED — E_REJUV CALIBRATED (84.7 J, 12 Pages Target)
Cites PAPER1: 5 | Cites PAPER3: 5