Navigating the Energy Doldrums: Can We Exploit Energy-Price Volatility To Lower the Cost of Computing?

Abstract

Energy costs are a major factor in the total cost of ownership (TCO) for high-performance computing (HPC) systems. The rise of intermittent green energy sources and reduced reliance on fossil fuels have introduced volatility into electricity markets, complicating energy budgeting. This paper explores variable capacity as a strategy for managing HPC energy costs -- dynamically adjusting compute resources in response to fluctuating electricity prices. While this approach can lower energy expenses, it risks underutilizing costly hardware. To evaluate this trade-off, we present a simple model that helps operators estimate the TCO impact of variable capacity strategies using key system parameters. We apply this model to real data from a university HPC cluster and assess how different scenarios could affect the cost-effectiveness of this approach in the future.

Figures (7)

• Figure 1: Electricity production and spot-market price in Germany over an average day in 2024. Data source: SMARD SMARD
• Figure 2: Visualization of our model for energy prices, demonstrated using Germany's historic day-ahead prices from 2024 (data source: SMARD SMARD, resolution: 1 hour). The left diagram plots the prices chronologically, while the right one depicts the price samples in descending order over a logarithmic x-axis, akin to a survival function from statistics. For a given shutdown fraction $x$ (e.g., $x = 1.15\%$), the price threshold is determined as defined in \ref{['eq:threshold']}, which is then used to categorize the price samples into regions of low and high prices, illustrated by the blue and orange areas.
• Figure 3: Price variability $PV$ of Germany's spot-market electricity prices from 2024 SMARD, for different sampling intervals, as defined in \ref{['fig:pv']}. The blue area depicts the range of $k$ for which our model assesses temporary shutdowns to be viable, based on our model in \ref{['eq:model']} and our estimation for Lichtenberg's $\Psi_{LB} = 2$. The point where a k-x-line leaves the blue area represents the point after which shutdowns are no longer beneficial (break-even point $x_{BE}$). For weekly samples, the model predicts that shutdowns are counterproductive in every case. For $PV_{1h}$, the model predicts that shutdowns are beneficial when $x < 3.32\%$. Note that both axes are logarithmic.
• Figure 4: Comparison of the respective price variabilities $PV$ of hourly intra-day spot-market prices (Australia: dispatch price) of Germany SMARD and South Australia AEMO. Similar to \ref{['fig:kx']}, the orange and blue areas represent the models viability prediction for $\Psi = 2$ and the hatched lines mark the respective break-even points after which shutdowns are no longer beneficial. Both axes are logarithmic.
• Figure 5: Maximum theoretical CPC reduction of temporary shutdowns over a no-shutdown policy for varying values for the cost-distribution coefficient $\Psi$, assuming the Germany's 2024 historic prices (resolution: 1 hour). The x-axis is logarithmic.