Power Usage Estimation of a RISC-V Platform using RTOS Events

• Typ der Arbeit: Bachelorarbeit
• Status der Arbeit: in Planung
• Projekte: AHA
• Betreuer: Tim-Marek Thomas, Daniel Lohmann

Context

Within the AHA project, we focus on optimizing embedded systems by specialization of the OS itself. The aim is to improve non-functional system properties like delay or memory footprint, which can reduce hardware costs and energy consumption.

For online optimization, e.g. via different scheduling approaches, and profiling we need to estimate the power usage. In the past, this was primarily done using a modelling approach via performance monitoring counters(PMCs)¹ or via special hardware components.

Problem

In a previous bachelor thesis(see below) an evaluation setup was already created, which collects PMCs and power measurements during representative workloads. For this we provided the SoC³ and a measurement device (a Rohde & Schwarz™ HMO3004 digital oscilloscope). Based on those measurements several models were trained, which allowed an estimation of the power consumption.

However, a lot of microcontrollers feature only a limited set of PMCs (e.g. the provided SoC has only 2 PMC registers available), thus the accuracy of the estimation is limited. One way to solve this problem is to employ multiplexing of the PMC register ². Another approach is to utilize software defined "PMC"s to broaden the simultaneousily recordable data.

Goal

In this thesis, the setup shall be extended to use hooks provided by an operating system, in our case the tracing subsystem of Zephyr⁴. OS events shall then be used instead of hardware PMCs (or in conjunction), like the amount of context switches, access to resources like queues, etc. as model inputs. A short suggestion of suitable events is already proposed in the previous thesis.

In particular, the following steps need to be taken:

• Literature research regarding prior usage of (RT)OS events (or similar, software defined events) for power modeling
• Recreate the setup of the previous bachelor thesis
• Adjust the existing setup for collection of chosen OS events and current readings ⁵
• Evaluate the correlation of the collected events with the power consumption/current
• Implement and evaluate several models, which are based on a mixture of events and PMC event counters (or even only RTOS events)
• Optional: Extend the setup to a whole system approach, based on trickle-down effects⁶

Topics: C/C++, Python, RISC-V, Zephyr

References