Enabling users to work sustainably on shared institute computing resources

The VISPA project promotes sustainable computing on a mid-scale physics cluster by implementing user-centric tools—such as per-job energy monitoring, "green-window" scheduling based on renewable energy availability, and voluntary carbon footprint reporting—to foster environmental awareness and reduce emissions through behavioral change.

The Gist

The "Green Kitchen" Approach to Supercomputing: A Simple Guide

Imagine you are part of a massive, shared community kitchen. This kitchen is huge, filled with high-tech ovens, blenders, and stovetops. Everyone in the neighborhood—students, professional chefs, and hobbyists—uses it to cook their meals (which, in this paper, are "scientific data analyses").

The problem? This kitchen is located in an old, drafty building from the 1970s. You can’t easily install new insulation or a better heating system. Because of this, the people running the kitchen realized they couldn't fix the building to save energy; instead, they had to change how the cooks behave.

This paper describes how the VISPA project (a computing cluster at a German university) turned their "kitchen" into a sustainable, eco-friendly workspace using three clever strategies.

---

1. The "Smart Recipe" Guide (Better Resource Requests)

The Problem: Imagine a cook who needs to boil one egg, but they reserve four massive, industrial-sized pots and a giant burner to do it, just "to be safe." While they are using those giant pots, no one else can use them, even though most of the heat is being wasted.

The Solution: The researchers created a tool that looks at a cook's history. Before you start a new recipe, the system whispers, "Hey, last time you made this pasta, you only actually used one pot and half the burner. Why not request a smaller setup this time?" By asking for only what they actually need, more people can cook at the same time, and less energy is wasted heating up empty space.

2. The "Smart Light Switch" (Dynamic Machine Shutdown)

The Problem: In a giant kitchen, it’s common to leave all the ovens preheating and all the lights on, even if no one is actually cooking anything. This "idle" energy is like leaving a car engine running in the driveway all night.

The Solution: The researchers implemented a system that acts like a motion-sensor light. If no one has been in a specific section of the kitchen for five minutes, the system "powers down" the equipment in that area. As soon as someone walks in to start cooking, the equipment wakes back up. This prevents the "engine" from idling all day and night.

3. The "Green Cooking Hours" (Sustainability-Aware Scheduling)

The Problem: Not all electricity is created equal. Sometimes the power grid is running on "dirty" energy (like coal), and sometimes it’s running on "clean" energy (like wind and sun).

The Solution: Think of this like a "Green Cooking Hour." The system uses a traffic light:

• Red Light: The grid is using mostly fossil fuels.
• Green Light: The wind is blowing and the sun is shining, so the grid is full of clean energy.

The researchers gave users a choice: "You can start your meal right now (Red Light), or you can tell us to wait until the Green Light appears." If a user has a massive task that isn't urgent (like a slow-cooking stew), they can tell the computer, "Wait until the sun is out to cook this." This ensures the "cooking" happens when the energy is cleanest.

---

The "Digital Twin" (The Practice Kitchen)

Before they tried these changes on the real, expensive equipment, the researchers built a "Digital Twin."

Imagine building a perfect, virtual replica of the kitchen in a video game. They ran thousands of "fake" cooks through this virtual kitchen to see what would happen. They discovered that if you combine all three tricks—asking for smaller pots, turning off idle ovens, and waiting for the green light—you get a kitchen that is incredibly efficient, uses much less power, and produces much less pollution, all without making the cooks wait too long for their meals.

The Big Picture

The goal isn't to force scientists to work less; it's to give them a "dashboard" (called PETRA) so they can see their own "carbon footprint." It’s about moving from a culture of "Just give me all the power possible!" to a culture of "Let's use exactly what we need, when the planet can afford it."

Technical Summary

Technical Summary: Enabling Users to Work Sustainably on Shared Institute Computing Resources

Problem Statement

Institutional computing clusters, such as the VISPA project at RWTH Aachen University, often face significant sustainability challenges. Unlike modern high-performance data centers with optimized Power Usage Effectiveness (PUE), departmental clusters are frequently housed in older buildings (e.g., 1970s infrastructure) with limited options for hardware-level energy efficiency or advanced cooling retrofits. Consequently, the primary opportunity for reducing greenhouse-gas (GHG) emissions lies not in structural engineering, but in user-centric behavioral shifts and intelligent job orchestration. The core challenge is to increase resource awareness among a diverse user base (students and researchers) and provide tools to balance scientific productivity with environmental responsibility.

Methodology

The authors implemented a multi-layered approach combining real-time monitoring, user-facing communication tools, and a sophisticated simulation framework.

1. Energy Monitoring & Estimation:
   • The system uses metered Power Distribution Units (PDUs) to measure total node energy ($E_{machine}$).
   • It combines device-level telemetry (RAPL for CPUs, NVML for GPUs) with PDU data to estimate energy per individual job ($E_i$).
   • The estimation model accounts for three components: GPU energy ($E_{gpu,i}$), CPU energy ($E_{cpu,i}$), and a residual overhead ($E_{overhead,i}$) covering fans, memory, and interconnects.
2. Sustainability-Aware Scheduling:
   • The cluster integrates real-time renewable energy data from the Fraunhofer Institute ISE, using a "traffic-light" system (Green/Yellow/Red).
   • Users can opt-in to "green-window" scheduling via HTCondor, allowing jobs to defer execution until the renewable share of the German power grid is high, subject to a user-defined maxwait constraint.
3. User Awareness Tools (PETRA):
   • A communication platform (PETRA) provides users with command-line and chatbot interfaces to query their personal energy footprints, project-level aggregate consumption, and daily renewable energy forecasts.
4. Resource Optimization:
   • To prevent "oversubscription" (where users request excessive memory to avoid job failure), the system provides a terminal prompt that displays the actual maximum memory usage of similar past jobs, encouraging "right-sized" resource requests.
5. Digital Twin Simulation:
   • The authors developed a "Mini-VISPA" simulation framework using Dockerized HTCondor instances.
   • A deep neural network (5-layer fully connected architecture) was trained on historical production data to predict instantaneous power draw from resource occupancy (CPU/GPU utilization), achieving a 94% Pearson correlation.

Key Contributions

• Per-Job Energy Accounting: A methodology to decompose and attribute energy consumption to specific users and research projects in a shared environment.
• Voluntary Sustainability Framework: A non-coercive system where users choose between immediate execution and "green" scheduling, preventing the imposition of hard constraints on scientific workflows.
• Data-Driven Resource Requesting: A tool to mitigate the inefficiency of over-provisioned memory requests by leveraging historical job telemetry.
• Predictive Simulation Framework: A high-fidelity digital twin capable of evaluating the trade-offs between energy savings, $CO_2e$ reduction, and user-perceived quality of service (waiting times).

Results

The paper evaluates the impact of three measures: (A) Better Resource Requests, (B) Dynamic Machine Shutdown, and (C) Sustainability-Aware Scheduling.

• Scenario 1: Academic/Research-Intensive Workload (High Load):
  • The combination of all three measures (A+B+C) yielded the lowest total energy and $CO_2e$ emissions while keeping median waiting times below the baseline.
  • Measure A (better requests) significantly reduced waiting times by allowing denser job packing.
  • Measure C (green scheduling) reduced $CO_2e$ by shifting loads to high-renewable periods.
• Scenario 2: Realistic/Underutilized Workload (Low Load):
  • Dynamic Machine Shutdown (B) was the most effective lever, reducing power and $CO_2e$ by approximately 90% by eliminating idle power consumption.
  • Measure A (better requests) actually increased waiting times in this specific simulation due to fragmentation artifacts, suggesting that in low-utilization regimes, capacity management (shutdowns) is more critical than packing efficiency.

Significance

This research demonstrates that for mid-scale, departmental computing resources, technological interventions must be paired with behavioral empowerment. By making the "invisible" cost of computation (energy and carbon) visible and actionable, the VISPA project provides a scalable template for academic institutions to transition toward sustainable scientific computing without compromising the flexibility required for research.