Utilization Score Metrics

• Utilization Score is a collection of metrics that quantify how effectively a resource is used, clearly distinguishing raw availability from genuine operational performance.
• In cloud and network systems, scores based on resource ratios (e.g., available MIPS, RAM, and bandwidth) help optimize load balancing and reduce processing times.
• Across multimodal learning and schedulability tests, utilization scores serve as diagnostic tools to inform counterfactual analyses and improve system efficiency.

Utilization score denotes a family of metrics that quantify how effectively an available resource, information source, or operating budget is actually used. In the cited literature, the term does not denote a single standardized formula. It refers variously to capacity ratios, workload–machine fitness indices, conditional performance-drop measures, confidence-based attribution scores, and utilization-based schedulability tests. Across these formulations, the common objective is to distinguish mere availability from effective use, whether the object being used is compute capacity, network bandwidth, retrieved evidence, supporting context, cross-modal information, or feedback (Hafner et al., 2023, Luciano et al., 2020, Xia, 4 Jun 2026).

1. Conceptual scope and recurring formulations

The literature uses utilization scores in several mathematically distinct ways. Some formulations measure fraction of capacity consumed; others measure distance from a target operating region; others quantify performance loss under counterfactual removal of a modality, context, or claim; and some summarize multi-dimensional capability by averaging multiple sub-scores.

Setting	Object being utilized	Score construction
Multi-modal urban mapping	SAR or optical information	Conditional Utilization Rates from F1 degradation under branch cut-off
Cloud load balancing	VM free capacity	$$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}$$
WISE	Resource operation relative to target	Aggregated transformed $z$-scores with penalties
Evidence utilization in LLMs	Oracle-reference evidence advantage	ONCU across no-evidence, full/retrieved, and oracle conditions
SCORE	Retrieved atomic claims	Leave-one-out confidence drops averaged over claims

These constructions appear in substantially different research programs, but they share an epistemic role: they are intended to diagnose whether observed performance arises from the resource under study, from alternative resources, or from baseline behavior that would have occurred even without the purported input. That distinction is explicit in modality cut-off experiments, context perturbation protocols, oracle-reference normalization, and target-based system indices (Hafner et al., 2023, Luciano et al., 2020, Shomee et al., 10 Feb 2026, Xia, 4 Jun 2026).

2. Capacity-oriented scores in cloud, datacenter, and network systems

In cloud load balancing, the utilization score can be a direct resource-availability score used for online placement. The Score-Based Dynamic Load Balancer filters out VMs whose active tasks exceed a threshold, retrieves availableMIPS, availableRAM, and availableBW, checks resource sufficiency after min-max normalization of task demands, and assigns the task to the VM with the highest score,

$$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$

On CloudSim 7G, this score-based policy outperformed the throttled strategy, improving average response times by 34% and 37% in different scenarios, reducing data center processing times by an average of 13%, and decreasing operational costs by 15% over a 24-hour simulation (Sakib et al., 7 Aug 2025).

A related but distinct formulation appears in Flex, where the central quantity is cluster utilization rather than per-VM suitability. Flex defines node utilization as $L_i/C$ and cluster utilization as

$\frac{\sum_{i\in N} L_i}{N\times C},$

with scheduling decisions based on estimated usage rather than user requests alone. Evaluations based on Google cluster traces showed up to 1.74x more requests admitted and 1.6x higher utilization than traditional schedulers while maintaining QoS above 99% (Le et al., 2020). In this setting, the utilization measure is not only descriptive; it is the optimization target of an online controller that adjusts aggressiveness through QoS feedback.

Internet interconnection studies use a still more literal utilization score: observed traffic normalized by provisioned capacity. DeepField Networks’ IPFIX-based method aggregates sampled flow records into five-minute intervals and computes utilization as total bits transferred during the interval divided by provisioned capacity times interval duration. In a five-month dataset spanning seven ISPs and about 97% of their paid peering, settlement-free peering, and ISP-paid transit links, aggregate peak utilization across interconnects was roughly 50%, fewer than 4% of link aggregation groups exceeded 95% utilization in any five-minute interval, and less than 10% of overall capacity ever exceeded 95% utilization at peak (Feamster, 2016). Here, the score is explicitly a capacity fraction, but interpretation is constrained by aggregation and sampling.

3. Utilization as dependence in multimodal learning and GPU execution

In multi-modal urban mapping, utilization is operationalized as a branch’s dependence on information from the other modality. The dual-branch architecture uses a U-Net backbone for Sentinel-1 SAR input $x_{\rm sar}$ and a U-Net backbone for Sentinel-2 optical input $x_{\rm opt}$, connected by Multi-Modal Transfer Modules. Final predictions are fused as

$p=\frac{1}{2}(p_{\rm sar}+p_{\rm opt}).$

To estimate dependence, the model replaces the incoming feature vector from the opposite branch with its training-set average, thereby cutting off true cross-modal information. The paper then defines Conditional Utilization Rates using an accuracy metric $A(y,p)$, with F1 as the main metric. The reported results were F1 $0.613\pm0.006$ for the SAR branch, $z$0 for the optical branch, and $z$1 for fusion. The corresponding CUR values were $z$2, $z$3, and $z$4, leading the paper to conclude that optical data was under-utilized despite good fusion performance (Hafner et al., 2023).

This counterfactual style of utilization measurement differs sharply from coarse aggregate metrics. The GPU profiling study argues that overall GPU utilization and achieved occupancy are surface-level quantities that fail to characterize contention under colocation. It proposes profiling at the level of block schedulers, IPC, pipeline utilization, L1/L2 cache hit rates, and memory bandwidth, with slowdown computed as

$z$5

The study reports bandwidth utilization up to 90% with slowdowns approaching 2x, and for PyTorch torch.mm it observed a 1.7x latency increase under colocation (Elvinger et al., 28 Jan 2025). The methodological implication is that a single utilization number can conceal substantial resource-specific underuse or interference.

4. Context, evidence, and tool utilization in language systems

Document-level machine translation uses utilization scores to ask whether supporting context is actually used rather than merely correlated with correct output. One method is perturbation-based: compare BLEU, COMET, or CXMI when the model receives correct context versus random context. Another is phenomenon-specific attribution, where ALTI+ is used to estimate the percentage of an output token’s attribution that comes from annotated supporting context:

$z$6

On pronoun-focused evaluations, sentence-level models showed zero or negligible attribution to context, single-encoder concatenation models attributed 74–89% of pronoun prediction to supporting context, and the multi-encoder model showed very low attribution to context (Mohammed et al., 2024).

For long-context and retrieval-augmented LLMs, the four-condition protocol separates no-evidence answerability, full-context utilization, retrieval-conditioned utilization, and oracle-reference recoverability. Its central estimator is ONCU:

$z$7

with a clipped version constrained to $z$8 and reported only for denominator-valid groups where $z$9. Across Controlled-ONCU-safe16K, HotpotQA-ONCU, and 2WikiMultiHopQA-ONCU, using 18,000 ONCU-compatible predictions, the main empirical finding was a task-dependent bottleneck split: controlled synthetic settings primarily exposed full-context utilization failures, whereas realistic multi-hop settings primarily exposed retrieval-chain coverage failures (Xia, 4 Jun 2026). The score is therefore explicitly diagnostic rather than leaderboard-oriented.

The SCORE framework defines context utilization at the claim level. After extracting unique atomic claims $$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$0 from retrieved documents, it computes the contribution of each claim by leave-one-out confidence subtraction:

$$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$1

and aggregates these as

$$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$2

Across the reported settings, removing a single claim reduced answer confidence by about 1–2% on average, with some claims causing much larger drops of about 9–10%; human–automated alignment for context utilization was 0.43, while inter-annotator agreement was 0.35 (Shomee et al., 10 Feb 2026). This formulation treats utilization as causal contribution of retrieved evidence to answer confidence.

Tool-utilization benchmarks generalize the concept further. UltraTool defines comprehensive tool utilization as the ability to plan, create, and use tools in complex real-world tasks. It evaluates six dimensions—planning; tool-creation awareness and creation; tool-usage awareness, selection, and usage—and reports an overall score as the average across these evaluation metrics. The benchmark contains 5,824 samples across 22 domains and 2,032 tools, with an average of 12.27 steps and 2.74 tool calls per sample; GPT-4 achieved the highest overall score, 76.04% in Chinese and 74.58% in English (Huang et al., 2024). In this setting, utilization score becomes a composite capability measure rather than a resource-ratio or counterfactual-dependence estimator.

5. Composite indices, useful-work fractions, and utilization-based schedulability

A composite utilization score can also be constructed from deviation from desired operating targets. The WISE framework defines, for each resource $$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$3, a target utilization $$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$4, acceptable deviation $$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$5, and upper limit $$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$6. The normalized deviation is

$$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$7

which is transformed either by $$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$8 or by $$\text{Score}=\text{availableMIPS}+\text{availableRAM}+\text{availableBW}.$$9, and then aggregated with optional weights and hard-limit penalties. Four variants are defined: two tanh-based scores $L_i/C$0 and two exp-based scores $L_i/C$1. In the tanh variants, lower is better; in the exp variants, higher is better. The framework was evaluated on MongoDB and streaming workloads on AWS EC2 VMs, where all four scoring variants worked well, precision was often 1.0, and rank-biased overlap values were typically 0.6–0.9 (Luciano et al., 2020). WISE formalizes utilization as closeness to a target zone rather than raw consumption.

Distributed stream processing uses utilization in yet another rigorous sense: the fraction of total time available for useful work. With checkpoint interval $L_i/C$2, checkpoint cost $L_i/C$3, failure rate $L_i/C$4, failure detection and restart cost $L_i/C$5, topology depth $L_i/C$6, and message delay $L_i/C$7, the basic no-failure definition is

$L_i/C$8

The full model incorporates failures and DAG topology and yields a closed-form utilization expression for stream processing systems, together with an optimal checkpoint interval

$L_i/C$9

which the paper notes depends only on checkpoint cost and failure rate (Jayasekara et al., 2019). Apache Flink experiments reported utilization improvements for every case, especially as system size increased, with up to 126% improvement for high failure rates.

Real-time scheduling theory uses utilization score in an admission-test sense. The k2U framework starts from a $\frac{\sum_{i\in N} L_i}{N\times C},$0-point effective schedulability test and derives sufficient utilization-based tests under fixed-priority scheduling. Its core hyperbolic bound is

$\frac{\sum_{i\in N} L_i}{N\times C},$1

with related total-utilization and logarithmic bounds. For implicit-deadline tasks on uniprocessor rate-monotonic scheduling, the classical hyperbolic and Liu–Layland-style bounds emerge as special cases; the paper also reports a speed-up factor of 1.76322 for deadline-monotonic scheduling and a capacity augmentation factor of 3.62 for DAG tasks under the derived analyses (Chen et al., 2015). Here, utilization score is not observational but analytical: it is a sufficient condition for schedulability.

6. Interpretation, proxies, and methodological limits

A central theme across the literature is that utilization must be distinguished from raw success. The urban-mapping study showed good fusion performance together with under-utilization of optical data (Hafner et al., 2023). The MT work argued that BLEU alone is not informative about the benefits of context (Mohammed et al., 2024). The ONCU protocol stated that final-answer accuracy, retrieval recall, and citation overlap do not by themselves identify whether evidence was used (Xia, 4 Jun 2026). The GPU study likewise argued that occupancy and overall GPU utilization are insufficient to predict interference under colocation (Elvinger et al., 28 Jan 2025). These results converge on a methodological point: a utilization score is often introduced precisely because conventional output metrics conflate availability, baseline competence, and actual use.

Another recurring issue is the use of operational proxies. In the study of auto-grader feedback, utilization was operationalized as a navigation event to the unique web page hosting feedback for a submission. The study explicitly noted that it could not guarantee the student truly read or carefully engaged with the feedback, but used feedback-page visits as a proxy. Within that operational definition, nFeedbackChecks correlated positively with ProjectScore at 0.273 with $\frac{\sum_{i\in N} L_i}{N\times C},$2, and the probability that the next submission improved was 38.46% after a feedback check versus 33.77% without a check, with Fisher’s Exact Test $\frac{\sum_{i\in N} L_i}{N\times C},$3 (Zhang et al., 17 Jul 2025). This illustrates both the value and the limitation of utilization scores: they can reveal statistically meaningful associations while remaining contingent on the fidelity of the proxy.

Taken together, these studies show that utilization score is best understood as a diagnostic construct rather than a universal metric. Its exact meaning depends on what is being utilized, what counterfactual or target is adopted, and whether the score is intended for causal diagnosis, online control, capacity planning, or formal schedulability analysis. The most robust formulations make those choices explicit, separate availability from use, and report companion metrics when a single score would obscure the failure mode (Luciano et al., 2020, Xia, 4 Jun 2026).