Access Official Technical PDF

Share Summary:LinkedInEmail

Heterogeneous processor power modeling can help engineers evaluate the energy and timing effects of real-time workloads before changing scheduling policies or deploying them on embedded hardware. A model based only on CPU type and frequency may miss important differences: tasks with distinct instruction mixes and memory-access patterns can draw different power and respond differently to frequency changes.

That matters wherever embedded systems must balance energy use with timing requirements. For industrial automation teams assessing real-time computing platforms, simulation can provide a way to compare scheduling approaches while accounting for more realistic processor behavior. The approach described here extends a real-time scheduling simulator with workload-dependent power and execution-time models, calibrated against measurements from a heterogeneous Arm platform.

Why CPU frequency is not enough

Dynamic voltage and frequency scaling (DVFS) changes processor operating conditions to manage the trade-off between performance and energy consumption. A common simplified assumption is that power follows a quadratic relationship with frequency, while a task’s execution time scales inversely with frequency. These relationships can be useful abstractions, but they do not capture every workload or hardware effect.

Measurements on an Odroid-XU3 board with an Arm big.LITTLE processor showed a meaningful difference in power between workloads running at the same frequency. At maximum frequency on the big CPU cluster, the encryption workload measured 38% more power than a memory-intensive workload designed to cause frequent cache misses. The latter spends time stalled while waiting for memory access, rather than continuously performing computation.

This illustrates why a task’s behavior matters as well as the processor’s frequency. Real applications sit between such contrasting cases, depending on their instruction mix and data-access patterns. The measurements also found that power in the clock-gating idle state varied with frequency, so idle consumption is not necessarily represented by one fixed value.

A workload-dependent power model

The proposed model represents two broad contributors to CPU power: switching activity and leakage. Switching power is associated with transistor activity, including the charging of transistor gate capacitances and brief losses during switching. Leakage power accounts for current that flows even when transistors are not switching.

In the model, workload type affects the activity level used to represent switching. This makes it possible for different workloads to produce different estimated power at the same frequency and voltage. Leakage is also included, with its relationship to switching power simplified under the assumption of stable temperature. The resulting model is a practical “gray box”: it reflects physical behavior without requiring a detailed description of processor internals.

Rather than relying on a single set of parameters for every task, the model is fitted to measurements for different workload types. The fitting process uses benchmark data from a real platform. That empirical step is important: a model’s usefulness depends not only on its equations, but also on how well its parameters represent the hardware and workloads being simulated.

Execution time needs workload context too

Power is only part of the energy picture. Energy use during a computation also depends on how long the task runs, and execution time can vary with both CPU frequency and workload characteristics.

A simple inverse-frequency rule assumes that all computation speeds up in direct proportion to frequency. In practice, that assumption can break down when a task encounters bottlenecks that do not scale in the same way, such as memory-access delays. To represent this behavior, the proposed execution-time model combines three elements:

  • A fixed offset for bottlenecks whose speed is not dependent on CPU frequency.
  • An inverse-frequency component representing idealized scaling of computation time.
  • An adjustment to the curve’s slope to better fit observed workload behavior.

This gives the simulator a way to represent workload-specific timing across CPU frequencies rather than applying one universal scaling rule. The model is intended for single tasks running on the system; it does not account for interference from other tasks, cache or bus contention, or DMA activity on the bus. Those effects would require additional modeling.

Integrating the models into real-time simulation

The models were implemented in RTSIM, an open-source simulator for real-time systems. RTSIM represents task execution, scheduling policies, CPUs, and system events. The extensions allow a task to identify its workload type and provide a computation duration calibrated at the maximum frequency of the highest-performance CPU. The simulator then adjusts execution time and power estimates according to the workload, CPU characteristics, and selected frequency.

A power-tracing component can track estimated consumption for each CPU during a simulation. This lets users examine how scheduling decisions affect both timing and power, rather than treating energy as separate from task execution. For system integrators and technical decision-makers, that can support preliminary comparisons of energy-aware scheduling strategies and their ability to meet timing constraints.

What validation showed—and what it does not

The models were compared with measurements from an Odroid-XU3 board containing four Cortex-A7 and four Cortex-A15 cores. Power measurements came from the board’s INA231 power meters. The tested workloads included compression, encryption, checksumming, and a cache-intensive benchmark, as well as idle behavior.

The simulator reproduced the measured power trends for the tested workloads, though a noticeable error appeared for the LITTLE CPU at mid-range frequencies. The authors suggested that the CPU power supply could contribute to this discontinuous behavior. Across the reported comparisons, the highest relative least-squares error for power was 16.1%. For execution time, the highest reported error was 3.57%. These results support the model as a useful simulation approach for the tested platform and workloads, not as a guarantee of accuracy on other processors or applications.

Further work identified by the authors includes improving representation of mixed and parallel workloads, and investigating models suited to general CPU-bound or memory-bound behavior. In practical use, model parameters should therefore be grounded in measurements relevant to the target hardware and workload classes.

Practical implications

For teams evaluating real-time computing platforms, the key lesson is that frequency alone may not tell the full story. Workload-aware power and execution-time estimates can give scheduling simulations a more representative basis for comparing energy and timing outcomes. The approach is most useful as a calibrated engineering tool: it can inform early evaluation, while its assumptions and coverage should be checked against the intended hardware, workloads, and system-level interference.

Access Official Technical PDF

About The Author