Research map / Architectures and efficiency / RNNs and normalization
Optimizing IT FinOps and Sustainability through Unsupervised Workload Characterization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a fully unsupervised framework that clusters VM utilization time series with M-TICC, maps clusters to inactive/active phases, and classifies VMs as non-productive, constantly productive, or alternating to produce shutdown or parking timetables.
Problem
Enterprise VMs often sit idle or unnecessarily provisioned, wasting cost and energy; existing threshold-based idle detection is overly conservative and supervised alternatives need costly labeled data.
Approach
Unsupervised subsequence clustering (M-TICC) of multivariate utilization metrics, abstracted into inactive and active phases via majority voting over per-metric cluster medians, then used to classify workloads and derive periodic parking timetables for alternating workloads.
Key results
- Classified 35.6% of VMs as non-productive, 27.2% alternating, and 37.2% constantly productive, versus only 1.1% non-productive by threshold-based methods.
- All VMs flagged by threshold-based methods were also flagged non-productive by the proposed approach.
- Of 41 VMs sent for feedback, 9 were decommissioned and 4 resized, with 18 deemed necessary by owners.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use data-driven unsupervised characterization instead of conservative threshold-based detection | Enterprise VM workload classification · Enterprise VMs | identified non-productive VMs: 1.1% -> 35.6% | Helped |
| Classify VMs using utilization patterns rather than threshold-based methods | Enterprise VM workload classification · Enterprise VMs | recall of threshold-identified VMs | Helped |
| Use proposed recommendations to guide VM actions | 41 VMs deemed non-productive · Enterprise VMs | owner-initiated decommissioning: 9 of the 41 | Helped |
| Use proposed recommendations to guide VM actions | 41 VMs deemed non-productive · Enterprise VMs | owner-initiated resizing: 4 of the 41 | Helped |
| Classify using resource utilization percentages | VMs deemed non-productive · Enterprise VMs | recommendation validity: 4 of the 41 | Mixed |
Surprising findings
- Non-productive VMs sometimes show high utilization during activities such as virus scans or system updates, so low utilization does not always equal idleness.
- Threshold-based idle detection flagged only 1.1% of VMs, far fewer than the unsupervised approach.
Limitations
- No ground-truth labels, so classification quality cannot be directly evaluated.
- Uses only resource utilization percentages, which can misread over-provisioned systems as inactive.
- Early-stage feedback limited to 41 VMs.
Experimental setup
Datasets: enterprise VM utilization metrics (VCPU util, VMEM util, NET, IO) from thousands of VMs
Metrics: percentage of VMs classified non-productive recall versus threshold-based methods number of inactive clusters application owner feedback on 41 VMs
Baselines: threshold-based idle VM detection (AWS CloudWatch, VMWare, Google) Duan et al. supervised binary classifier
Builds on: M-TICC Toeplitz inverse covariance clustering
Closest papers by meaning
- Spatial-Temporal Heterogenous Graph Contrastive Learning for Microservice Workload Prediction (AAAI 2025 · similarity 0.81)
- DCcluster-Opt: Benchmarking Dynamic Multi-Objective Optimization for Geo-Distributed Data Center Workloads (NeurIPS 2025 · similarity 0.81)
- FastPERT: Towards Fast Microservice Application Latency Prediction via Structural Inductive Bias over PERT Networks (AAAI 2025 · similarity 0.80)
- AI Progress Should Be Measured by Capability-Per-Resource, Not Scale Alone: A Framework for Gradient-Guided Resource Allocation in LLMs (NeurIPS 2025 · similarity 0.79)
- Carbon Footprint Reduction for Sustainable Data Centers in Real-Time (AAAI 2024 · similarity 0.79)
- SustainDC: Benchmarking for Sustainable Data Center Control (NeurIPS 2024 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.