SimBricks enables us to measure full-system performance numbers like end-to-end application throughput and latency for custom hardware designs, such as offloading computations from the CPU, in simulation. However, energy consumption is often an equally important metric. For example, we might be interested in the amount of energy we save compared to running our workload in software only. Furthermore, when designing this hardware, we want to make well-informed decisions for tuning design parameters that trade off speed or computational capabilities and energy consumption. Therefore, we need to be able to evaluate the design early in the process.
Considering that energy consumption is spread across all of a system’s components and also changes dynamically based on the workload, we need dynamic information to make an accurate energy consumption estimation. We can collect this during simulation without perturbing the workload.
Overall, we take a modular approach by estimating the energy consumption individually per component and afterwards summing up into one number for the complete system. To this end, we use SimBricks to run the unmodified application and actual workload. We instrument the full-system simulation to collect the information necessary for accurate energy estimation and draw on existing work for estimating the energy consumption of component simulators like GemStone and the power analysis tools from hardware tool chains. Herein, timing information from the simulation can be used to compute energy numbers from tools that only provide power estimates.