Research
SLED Lab builds system software for the layer where computing meets the physical world — processors, batteries, power, and heat — on edge devices from smartphones to electric vehicles. Today this software follows rules fixed at design time, blind to each device’s workload, user, and environment; we replace those rules with run-time learning, grounded in scheduling theory and domain expertise. Our results range from DVFS governors that learn their thermal surroundings (ICCAD 2025) to smartphone batteries that cut users’ low-battery time (MobiSys 2023) and schedulers that more than double battery lifespan (RTSS 2019); we are now bringing the same principles to AI agents and LLM serving.
A. Learning-based system optimization
Keywords: reinforcement learning · on-device AI · healthcare
Rule-based system software — DVFS governors, thermal throttling — is tuned by vendors for universal use cases, so it cannot adapt to each app's workload or the device's thermal environment. EarDVFS (ICCAD 2025) is a reinforcement-learning DVFS governor that steers, rather than replaces, the vendor-tuned governor, improving power efficiency by 21.6% on average. The same learning-driven approach powers on-device applications, from contactless arrhythmia diagnosis with radar (mCardiacDx, JTEHM 2026) to real-time per-app energy prediction (Serenus, UIST 2024).
Representative:
EarDVFS: Environment-Adaptable RL-based DVFS… (ICCAD 2025)
mCardiacDx: Radar-Driven Contactless Monitoring… (JTEHM 2026)
Serenus: Alleviating Low-Battery Anxiety… (UIST 2024)
B. System-level support for battery & energy
Keywords: battery management systems · embedded systems · energy efficiency
A battery's real capability is set not only by its chemistry but by the system software that charges, discharges, and combines its cells. MixMax (MobiSys 2023) mixes three complementary battery types on a smartphone and co-optimizes their ratio and charge/discharge policies, cutting users' low-battery time by up to 24.6%. The same system-level approach extends to large-scale battery systems and EV infrastructure, from reconfiguration-assisted charging (TII 2024) to wait-time-guaranteed battery swap stations (RTAS 2025).
Representative:
MixMax: Leveraging Heterogeneous Batteries… (MobiSys 2023)
Leveraging Customized Heterogeneous Batteries… (TSUSC 2025)
Scheduling EV Battery Swap/Charge Operations (RTAS 2025)
RAC+: Supporting Reconfiguration-Assisted Charging… (TII 2024)
C. Novel resource management framework
Keywords: scheduling algorithms · AI agents · LLM serving
Scheduling shapes the physical behavior of a system, not just its deadlines. We showed that task scheduling systematically affects battery aging (RTSS 2019): our RET (Reserved Execution Time) framework keeps offline timing guarantees intact while runtime heuristics flatten the power draw, extending battery lifespan by up to 144.4%. We are now carrying these principles to AI-agent workloads: cache- and microarchitecture-aware CPU affinity management speeds up concurrently running AI agents (under submission), a first step toward power- and thermal-aware scheduling for LLM-serving systems.
Representative:
Battery Aging Deceleration… (RTSS 2019)
Non-Preemptive Real-Time Multiprocessor Scheduling… (RTSS 2020)
Battery-Aging-Aware Run-Time Slack Management… (JSA 2023)
