When a modern microprocessor runs a heavy workload, it generates heat. If a single core or region of the chip gets too hot, the hardware triggers Dynamic Thermal Management (DTM)—usually in the form of Dynamic Voltage and Frequency Scaling (DVFS)—to throttle the clock speed and protect the silicon. While effective at preventing physical damage, this reactive approach treats the processor as a uniform entity, dragging down performance and introducing severe synchronization delays in parallel applications.
Thermal-aware workload distribution shifts the paradigm from throttling clock speeds to orchestrating where and when computation occurs. This approach works because silicon chips possess two physical properties that traditional schedulers ignore: spatial thermal variation (different parts of a chip cool at different rates and bleed heat into one another) and thermal capacitance (silicon takes time to heat up). By strategically mapping, migrating, and intermixing tasks, systems can exploit these properties to maximize sustained performance within safe thermal boundaries.
Exploiting Spatial Topology and Lateral Heat Transfer
A processor is not an isolated collection of independent cores; it is a continuous physical package where heat moves laterally. Schedulers must account for spatial thermal coupling—the reality that a high-intensity workload on one core will bleed heat into adjacent idle cores, accelerating hardware degradation.
To mitigate this, Fatemeh Hossein Khani et al. developed a two-stage mapping strategy in A Two-Level Thermal Cycling-aware Task Mapping Technique for Reliability Management in Manycore Systems. The authors use density-based spatial clustering to group cores into "bins" of similar current temperatures, assigning new tasks to clusters that match their expected thermal profiles. By explicitly accounting for lateral heat transfer between neighbors, this approach prevents steep thermal gradients and thermal cycling, extending system lifespan with minimal performance overhead.
Furthermore, a chip's physical boundary dictates its cooling efficiency. Cores located at the perimeter of a die dissipate heat much faster than those trapped in the center. In Thermal Management for S-NUCA Many-Cores via Synchronous Thread Rotations, Sobhan Niknam et al. exploit this by partitioning cores into concentric rings based on their Manhattan distance from the chip's center. This topology-aware grouping allows the scheduler to manage threads across predictable thermal zones.
On a larger scale, this spatial awareness extends to the server chassis and data center. Niraj H. Tolia et al. demonstrate in Delivering Energy Proportionality with Non Energy-Proportional Systems - Optimizing the Ensemble that modeling the non-uniform spatial relationships between specific cooling fans and heat-generating components allows controllers to activate only the most effective cooling elements, optimizing the entire ensemble's energy footprint. Similarly, Meikang Qiu et al. note in Guest Editorial: Green and Cloud Computing that modeling the long-term thermal profiles of high-density storage clusters transforms heat management into a primary design constraint for cloud infrastructure.
Proactive Migration and Thread Rotation
Because silicon has thermal capacitance, a core does not instantly reach its peak temperature when a task begins. Schedulers can exploit this delay by moving workloads before a hotspot forms.
In Network-on-Chip (NoC) architectures, Vijaykrishnan Narayanan shows in Hotspot Prevention Through Runtime Reconfiguration in Network-On-Chip that periodically remapping computational tasks across a processing grid using simple algebraic transformations can lower peak temperatures by up to 8°C. This dynamic spatial redistribution proves far more effective than static thermal-aware placement.
Traditional DTM (Reactive Throttling):
[ Core 0: HOT ] ---> Trigger DVFS ---> [ Core 0: Throttled / Slow ]
Migration-Centric Management (Proactive Rotation):
[ Core 0: Warming ] --(Low-overhead Migration)--> [ Core 1: Cool ]
[ Core 0: Cooling ] <---------------------------- [ Core 1: Warming ]
Historically, the barrier to frequent task migration has been the performance penalty of moving thread contexts and reloading caches. However, modern hardware architectures change this trade-off. Sobhan Niknam et al. leverage Static Non-Uniform Cache Architectures (S-NUCA), where the last-level cache is logically shared but physically distributed. Because data remains accessible across the shared network, the overhead of a thread migration is exceptionally low. Their framework, HotPotato, uses synchronous thread rotations as the primary thermal management lever, maintaining high clock speeds by moving the workload faster than the silicon can overheat.
Temporal Intermixing and Heterogeneous Mapping
When tasks cannot be easily migrated, systems can manage heat by controlling the sequence and composition of workloads. As established by Marek Chrobak et al. in Algorithms for temperature-aware task scheduling in microprocessor systems, the specific sequence of task execution fundamentally dictates processor temperature and overall throughput, proving that competitive online scheduling algorithms can bound thermal behavior.
One powerful manifestation of this is workload intermixing. In Many-Core GPGPUs, Rajib Nath et al. introduced TABS in Temperature aware thread block scheduling in GPGPUs. GPGPUs feature massive data parallelism with a short lifetime for individual thread blocks, meaning new blocks can be scheduled with almost zero overhead. The authors' runtime scheduler dynamically intermixes "hot" (compute-bound) and "cold" (memory-bound) kernels on the same execution units, balancing the thermal load over time and preventing the system from ever hitting a throttling threshold.
Workload Intermixing Timeline:
| Compute (Hot) | Memory (Cold) | Compute (Hot) | Memory (Cold) |
^---------------------------------------------------------------^
Steady, Safe Temperature
In heterogeneous environments, this requires integrating thermal profiling directly into the application partitioning framework. Eduardo Weber Wächter et al. demonstrate in Reliable mapping and partitioning of performance-constrained openCL applications on CPU-GPU MPSoCs that explicitly profiling temperature alongside performance when mapping OpenCL applications onto CPU-GPU systems reduces average chip temperatures and boosts long-term reliability.
Similarly, for deep learning workloads on mobile devices, Sicong Liu et al. note in Deep Learning Inference on Heterogeneous Mobile Processors: Potentials and Pitfalls that distributing inference across multiple heterogeneous processors (CPUs, GPUs, NPUs) prevents any single component from reaching critical thermal thresholds, leveraging the combined heat dissipation capacity of the entire system-on-chip.
Reclaiming and Reallocating Thermal Headroom
Thermal management can also operate by treating the chip's total Thermal Design Power (TDP) as a shared, dynamic pool. When one part of a system slows down or goes idle, it leaves behind "stranded" thermal headroom that another component can immediately exploit.
In high-performance parallel computing (HPC), applications often suffer from load imbalances where some cores sit idle waiting for others to reach a synchronization barrier. Daniele Cesarini et al. developed COUNTDOWN in COUNTDOWN: A Run-Time Library for Performance-Neutral Energy Saving in MPI Applications to identify this "communication slack." By placing waiting MPI processes into deep sleep states, the runtime library drastically cools those specific cores. This sudden drop in localized heat frees up the overall package's thermal budget, allowing the remaining active cores to safely Turbo Boost to higher frequencies, often accelerating the entire application beyond its baseline.
Shared Thermal Budget Pool:
+-----------------------------------+
| Core 0 (Active) | Core 1 (Idle) | -> Core 1 Sleeps, cooling down
+-----------------------------------+
| Core 0 (TURBO) | Core 1 (SLEEP)| -> Core 0 absorbs the freed thermal budget
+-----------------------------------+
This dual-tier optimization—balancing long-term thermal horizons with short-term adjustments—is explored by Daniele Cesarini et al. in Modeling and Evaluation of Application-Aware Dynamic Thermal Control in HPC Nodes. By mapping critical parallel tasks to the naturally coldest physical cores and dynamically adjusting frequencies on a short-term horizon, the system delays thermal throttling across the entire parallel application.
This cross-domain reallocation works equally well across different hardware components. In mobile processors, Jawad Haj-Yahya et al. introduced SysScale in SysScale: Exploiting Multi-domain Dynamic Voltage and Frequency Scaling for Energy Efficient Mobile Processors. Instead of managing the CPU in a silo, SysScale dynamically scales down the voltage and frequency of underutilized I/O and memory domains, instantly reallocating that thermal headroom to the compute-heavy CPU or GPU.
Finally, in real-time and environmental systems, this budget management must adapt to external feedback loops. Kai Huang et al. propose Adaptive Periodic Thermal Management (APTM) in Adaptive Periodic Thermal Management for Pipelined Hard Real-Time Systems, which reclaims slack from early task completions in a multi-core pipeline to inject periodic sleep cycles into the hottest stages, balancing the thermal load dynamically. For edge devices, Noman Bashir et al. demonstrate in Jointly Managing Electrical and Thermal Energy in Solar- and Battery-powered Computer Systems that integrating environmental thermodynamic models with convex optimization frameworks—similar to the liquid-cooling optimization frameworks proposed by Giovanni De Micheli et al. in A combined sensor placement and convex optimization approach for thermal management in 3D-MPSoC with liquid cooling—allows systems to safely extract up to 11% more computation by co-managing electrical and thermal feedback loops.
By treating heat not as an emergency to be mitigated by stopping the clock, but as a predictable, fluid resource distributed across space and time, modern runtimes can keep processors running faster, longer, and more reliably.
Key Foundational Work
Power efficient resource scaling in partitioned architectures through dynamic heterogeneity (2006) — contributed 1 concept elements to this direction.
Improving Thermal-Safe Test Scheduling for Core-Based Systems-on-Chip Using Shift Frequency Scaling (2005) — contributed 1 concept elements to this direction.