🔬
Literature Positioning: Datacenter GPU (Cotroneo 2026) vs. Edge CPU (Our Thesis)
O estudo de Cotroneo & Cukic (2026) investigou envelhecimento em GPUs de nuvem (vLLM / Triton) e constatou vazamento em memória privada sem perda perceptível de vazão (TPS) em 36h. Em contraste, o nosso trabalho no Edge CPU demonstra que a contenção nos canais de memória DDR compartilhados e cache L3 provoca decaimento mensurável de vazão (-1.2% a -4.8%) e penalidade energética em Joules, justificando o rejuvenescimento ativo periódico calibrado em ΔErejuv = 84.7 J.
1. Estado dos 4 Artigos do Doutoramento (Thesis Track)
PAPER1-IOTJ
IEEE Internet of Things Journal (IoT-J)
Submitted & Under Peer Review (Tracking: IoT-77532-2026)
Slowing Down to Save: A Longitudinal, Carbon-Aware Evaluation of Large Language Model Inference on CPU-Only Edge Nodes
Covers 11 longitudinal 48h campaigns; establishes Pareto-optimal CPU frequency scaling; 100% Rule 5 compliant.
Size: 12 / 12 exact (0 overfull)
PAPER2-AGING-POSITION
IEEE Trans. on Emerging Topics in Computing (TETC) SI WoSAR 2026 (Q1, IF 4.8) | AISAR Synergy
Ready for Review (Camera-Ready Alignment)
Software Aging and Rejuvenation for Large Language Model Inference on IoT and Edge Nodes: A Position and Research Agenda
Defines Aging-Related Bugs (ARBs) in edge runtimes; 4-column literature taxonomy; 12/12 advisor comments resolved.
Size: 5 / 6 exact (0 overfull)
PAPER3-MECHANISM
IEEE/IFIP DSN 2027 Disrupt Track PER (Berlin) — dsn2027-berlin.github.io
Draft Complete (Exact 7 Pages, Zero Overfull)
Layer Localization and Mechanism Isolation of Performance Decay in Sustained Edge LLM Inference
3-layer isolation (L1 thermal, L2 page cache, L3 heap dirty); smaps_rollup proves 98.5% heap fraction; ARB-MEM-01 & ARB-CORO-01.
Size: 7 / 7 exact (0 overfull)
PAPER4-FRAMEWORK
IEEE Trans. Green Communications and Networking (TGCN) / IEEE TSE
Architecture & Optimal Scheduling Formulated
Energy- and Carbon-Aware Software Rejuvenation Scheduling for Edge LLM Services
Closed-form T* model balancing degradation carbon penalty against rejuvenation overhead (Delta E_rejuv = 84.7 J calibrated from Exp D).
Size: 12 target (Formulation Stage)
2. Pilares de Investigação (A1–A4) & Research Questions
RQ1-SEPARABILITY
Pillar A1 • paper2-aging-position / paper3-mechanism
How can progressive software aging in long-running LLM inference runtimes (e.g., Ollama / llama.cpp) be statistically isolated from thermal throttling and DVFS governor drift under variable arrival workloads?
Hipótese: Under fixed hardware frequency caps that eliminate turbo thermal saturation, any residual monotonic drift in per-token latency or RSS memory over >24h is attributable to runtime/OS state accumulation rather than thermal decay.
RQ2-LAYER-LOCALIZATION
Pillar A2 • paper3-mechanism
Which intervention (idle thermal pause, OS page-cache drop via POSIX_FADV_DONTNEED, or runtime warm restart) restores throughput and Joules-per-token efficiency without requiring a full node reboot?
Hipótese: In edge LLM inference, memory allocator fragmentation and runtime dirty heap accumulation account for the vast majority of non-thermal throughput drift, recoverable via lightweight process restart without OS reboot.
RQ3-JOULE-REJUVENATION
Pillar A3 • paper4-framework
When is the energy and carbon cost of a rejuvenation action (Delta E_rejuv) amortized by the post-rejuvenation reduction in Joules per token?
Hipótese: Energy-aware rejuvenation is net-positive in Joules when cumulative throughput degradation slope alpha exceeds sqrt(2 * Delta E_rejuv / (3600 * P_load * alpha)); co-scheduling with low-carbon grid windows minimizes gCO2eq.
RQ4-VERIFIABLE-PROTOCOL
Pillar A4 • paper2-aging-position / paper4-framework
What minimal continuous out-of-band telemetry invariants are required to certify that a multi-day edge AI experiment actually maintained its declared hardware and workload state?
Hipótese: Boundary-only configuration checks fail silently in >15% of multi-day edge campaigns; continuous in-band + out-of-band cross-auditing detects 100% of silent cap violations and sensor dropouts.
RQ5-WORKLOAD-NONSTATIONARITY
Pillar A1 • paper3-mechanism
Does bursty/diurnal prompt arrival accelerate KV-cache allocator fragmentation compared to closed-loop saturation at equal total token volume?
Hipótese: Variable prompt lengths and bursty inter-arrival times trigger frequent KV-cache reallocations, increasing heap fragmentation compared to uniform saturation.
RQ6-MOE-EXPERT-HOTSPOTS
Pillar A2 • paper3-mechanism
Do Mixture-of-Experts (MoE) models (e.g., DeepSeek-v2 Lite) exhibit different page-cache and RSS aging trajectories than dense models due to sparse expert routing?
Hipótese: Sparse activation across experts creates non-uniform page access patterns that interact differently with kernel memory management over multi-day runs.
RQ7-CARBON-AWARE-REJUVENATION-WINDOW
Pillar A3 • paper4-framework
Can rejuvenation actions (warm runtime reset) be co-scheduled with low-carbon grid windows (e.g., solar midday in PT/ES) to reduce net rejuvenation carbon cost by >40%?
Hipótese: Aligning proactive rejuvenation with diurnal grid carbon troughs minimizes the gCO2eq penalty of reload spikes while preventing late-stage aging degradation.