pith. sign in

arxiv: 2604.03145 · v1 · submitted 2026-04-03 · 💻 cs.NI · cs.DC

Causal Inference for Quantifying Noisy Neighbor Effects in Multi-Tenant Cloud Environments

Philipe S. Schiavo , Jo\~ao P. S. Milanezi , Mois\'es R. N. Ribeiro , V\'ictor M. G. Mart\'inez , Jo\~ao Henrique Corr\^ea , Jos\'e Marcos Nogueira , Fernando Frota Redigolo , Tereza C. Carvalho
show 1 more author
Fl\'avio de Oliveira Silva
This is my paper

Pith reviewed 2026-05-13 18:29 UTC · model grok-4.3

classification 💻 cs.NI cs.DC
keywords noisy neighborcausal inferenceGranger causalitymulti-tenant cloudperformance degradationKubernetesresource contention
0
0 comments X p. Extension

The pith

Controlled experiments paired with Granger causality quantify noisy neighbor effects in multi-tenant clouds and identify resource-specific degradation signatures.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a method that runs controlled workloads in a Kubernetes testbed and applies multi-stage causal inference to measure how one tenant interferes with others. It reports concrete slowdowns, such as 67 percent in I/O-bound cases under combined stress, and uses Granger causality to count a 75 percent rise in causal links once the noisy neighbor starts. The same analysis extracts distinct degradation patterns tied to CPU, memory, disk, or network contention. These patterns turn noisy-neighbor effects from an opaque problem into a measurable and diagnosable one. Cloud operators could therefore base SLA enforcement and resource scheduling on the observed causal structure rather than on anomaly alerts alone.

Core claim

The methodology combines controlled experimentation with multi-stage causal inference to quantify performance degradations up to 67 percent in I/O-bound workloads under combined stress and statistically establishes causality through Granger analysis, revealing a 75 percent increase in causal links when the noisy neighbor activates, while identifying unique degradation signatures for each resource contention vector.

What carries the argument

Multi-stage causal inference that applies Granger causality to time-series metrics collected from controlled experiments in a Kubernetes testbed.

If this is right

  • Performance degradations reach 67 percent in I/O-bound workloads under combined stress.
  • Activation of a noisy neighbor produces a 75 percent increase in detected causal links.
  • Each resource contention type (CPU, memory, disk, network) generates a distinct degradation signature.
  • The signatures enable diagnostic capabilities that extend beyond simple anomaly detection for SLA management.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Cloud schedulers could incorporate the signatures to isolate or throttle tenants before degradations breach SLAs.
  • The same experimental-plus-Granger pipeline could be adapted to quantify interference in shared edge or HPC environments.
  • Production deployments would benefit from adding tests for hidden confounders that the controlled testbed may miss.

Load-bearing premise

Results from the controlled Kubernetes testbed accurately represent real-world multi-tenant clouds and Granger causality on the collected metrics identifies genuine causal relationships without unmeasured confounding factors.

What would settle it

Replicate the same workload mixes on a public cloud provider and verify whether the degradation percentages and the 75 percent increase in causal links match the testbed observations.

Figures

Figures reproduced from arXiv: 2604.03145 by Fernando Frota Redigolo, Fl\'avio de Oliveira Silva, Jo\~ao Henrique Corr\^ea, Jo\~ao P. S. Milanezi, Jos\'e Marcos Nogueira, Mois\'es R. N. Ribeiro, Philipe S. Schiavo, Tereza C. Carvalho, V\'ictor M. G. Mart\'inez.

Figure 1
Figure 1. Figure 1: Impact heatmaps showing mean percentage change for victim tenants across experimental phases and metrics. Red indicates degradation; blue indicates [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Causal link density per phase. The noisy neighbor’s activation dramatically increases causal connections, proving directional influence. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Consolidated ECDFs showing unique distributional shifts for each noise type. Different contention vectors produce distinct ”degradation signatures.” [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Tenant Coupling Index across phases and metrics. The Noisy Neighbor ( [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
read the original abstract

Resource sharing in multi-tenant cloud environments enables cost efficiency but introduces the Noisy Neighbor problem, i.e., co-located workloads that unpredictably degrade each other's performance. Despite extensive research on detecting such effects, there are no explainable methodologies for quantifying the severity of impact and establishing causal relationships among tenants. We propose an analytical that combines controlled experimentation with multi-stage causal inference and validates it across 10 independent rounds in a Kubernetes testbed. Our methodology not only quantifies severe performance degradations (e.g., up to 67\% in I/O-bound workloads under combined stress) but also statistically establishes causality through Granger causality analysis, revealing a 75\% increase in causal links when the noisy neighbor activates. Furthermore, we identify unique "degradation signatures" for each resource contention vector (i.e., CPU, memory, disk, network), enabling diagnostic capabilities that go beyond anomaly detection. This work transforms the Noisy Neighbor from an elusive problem into a quantifiable, diagnosable phenomenon, providing cloud operators with actionable insights for SLA management and smart resource allocation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a methodology combining controlled experiments in a Kubernetes testbed with multi-stage causal inference to quantify noisy neighbor effects in multi-tenant clouds. It reports up to 67% performance degradation in I/O-bound workloads under combined stress and a 75% increase in causal links (via Granger causality) when the noisy neighbor activates, plus unique degradation signatures for CPU, memory, disk, and network contention vectors, validated over 10 independent rounds.

Significance. If the central claims hold after addressing causality limitations, the work would provide cloud operators with actionable diagnostic tools that go beyond anomaly detection, supporting better SLA management and resource allocation in shared environments. The degradation signatures represent a potentially useful contribution for distinguishing contention types.

major comments (3)
  1. [Abstract] Abstract: the quantitative claims of 67% degradation and 75% increase in causal links are presented without error bars, confidence intervals, p-values, or full statistical methods, undermining verifiability of the reported results from the 10 experimental rounds.
  2. [Causal inference] Causal inference section: Granger causality is applied to observed time-series metrics to establish the 75% increase in causal links, but the analysis does not include checks for unmeasured confounders (e.g., hypervisor scheduling, shared cache state, or OS noise), leaving open the possibility that reported causal relationships are spurious.
  3. [Experimental validation] Experimental validation: the assumption that the controlled Kubernetes testbed results generalize to real-world multi-tenant clouds is not supported by explicit tests for hidden variables or external validity, which is load-bearing for the claim of diagnosable degradation signatures.
minor comments (2)
  1. Clarify the exact multi-stage causal inference procedure, including how Granger tests are sequenced with other stages and any preprocessing of the time-series data.
  2. Add legends, axis labels, and statistical annotations to all figures showing degradation signatures and causal link counts.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the thoughtful and constructive comments. We address each major point below and indicate the planned revisions to improve the manuscript's clarity and rigor.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the quantitative claims of 67% degradation and 75% increase in causal links are presented without error bars, confidence intervals, p-values, or full statistical methods, undermining verifiability of the reported results from the 10 experimental rounds.

    Authors: We agree that the abstract should provide statistical context for the key quantitative results. In the revised version, we will include error bars, 95% confidence intervals, and p-values for the 67% degradation and 75% increase figures, computed across the 10 independent rounds. We will also briefly summarize the statistical procedures used (e.g., mean and standard deviation over rounds) to support verifiability while keeping the abstract concise. revision: yes

  2. Referee: [Causal inference] Causal inference section: Granger causality is applied to observed time-series metrics to establish the 75% increase in causal links, but the analysis does not include checks for unmeasured confounders (e.g., hypervisor scheduling, shared cache state, or OS noise), leaving open the possibility that reported causal relationships are spurious.

    Authors: Granger causality is known to be sensitive to unmeasured confounders, and we acknowledge this limitation in the controlled testbed setting. Our experiments used resource isolation, fixed allocations, and repeated independent runs to reduce external variability. In revision, we will expand the causal inference section with explicit stationarity tests, autocorrelation checks, and a sensitivity analysis (e.g., varying lag orders and adding synthetic noise). We will also add a limitations paragraph discussing potential residual confounders such as hypervisor effects. revision: partial

  3. Referee: [Experimental validation] Experimental validation: the assumption that the controlled Kubernetes testbed results generalize to real-world multi-tenant clouds is not supported by explicit tests for hidden variables or external validity, which is load-bearing for the claim of diagnosable degradation signatures.

    Authors: The manuscript presents results from a reproducible Kubernetes testbed chosen to enable controlled isolation of contention vectors. We agree that external validity to arbitrary production clouds is not directly demonstrated and will add an explicit limitations subsection acknowledging this. The degradation signatures are tied to observable resource metrics, which we argue provide a transferable diagnostic basis, but we will qualify the claims accordingly and outline future directions for production validation. revision: partial

standing simulated objections not resolved
  • Direct empirical tests of external validity in live production multi-tenant clouds, which would require access to operational systems outside the scope of the current controlled testbed study.

Circularity Check

0 steps flagged

No significant circularity; results derive from independent testbed measurements and standard statistical tests.

full rationale

The paper's core claims rest on controlled Kubernetes experiments run across 10 independent rounds, time-series metric collection, and application of standard Granger causality tests to quantify degradations (e.g., 67% I/O impact) and link increases (75%). These quantities are computed directly from observed data rather than fitted to the target claims by construction, renamed from prior results, or justified solely via self-citation chains. No equations or steps reduce the reported causality or signatures to the inputs by definition; the methodology remains externally falsifiable against the collected benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard assumptions from causal inference applied to experimental performance data; no new free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Granger causality tests applied to time-series performance metrics can establish causal relationships among co-located cloud workloads
    Invoked when the paper reports a 75% increase in causal links upon noisy neighbor activation.

pith-pipeline@v0.9.0 · 5553 in / 1317 out tokens · 40334 ms · 2026-05-13T18:29:22.619909+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

  1. [1]

    Multitenant containers as a service (caas) for clouds and edge clouds,

    B. C. S ¸enel, M. Mouchet, J. Cappos, T. Friedman, O. Fourmaux, and R. McGeer, “Multitenant containers as a service (caas) for clouds and edge clouds,”IEEE Access, vol. 11, pp. 144 574–144 601, 2023

    work page 2023

  2. [2]

    Detecting noisy neighbors in cpu-isolated cgroups environments,

    S. V olpert, S. Winkelhofer, J. Domaschka, and S. Wesner, “Detecting noisy neighbors in cpu-isolated cgroups environments,” inProceedings of the 16th ACM/SPEC International Conference on Performance Engi- neering, 2025, pp. 224–231

    work page 2025

  3. [3]

    Finding noisy neighbours and quantifying performance impact,

    M. Huynh, “Finding noisy neighbours and quantifying performance impact,”Proceedings of the 2020 OMI Seminars (PROMIS 2020), 2021, 2021

    work page 2020

  4. [4]

    Noisy Neighbor Influence in the Data Plane of Beyond 5G Networks,

    R. Moreira, L. F. Rodrigues Moreira, T. C. Carvalho, and F. de Oliveira Silva, “Noisy Neighbor Influence in the Data Plane of Beyond 5G Networks,” in2026 IEEE 23rd Consumer Communications & Networking Conference (CCNC), Jan. 2026, pp. 1–6, iSSN: 2331-9860. [Online]. Available: https://ieeexplore.ieee.org/document/11366571

    work page arXiv 2026

  5. [5]

    Noisy neighbour impact assessment and prevention in virtualized mobile networks,

    F. Muro, E. Baena, S. Fortes, L. Nielsen, and R. Barco, “Noisy neighbour impact assessment and prevention in virtualized mobile networks,”IEEE Transactions on Network and Service Management, vol. 20, no. 1, pp. 415–425, 2022

    work page 2022

  6. [6]

    An unsupervised approach to online noisy-neighbor detection in cloud data centers,

    T. Lorido-Botran, S. Huerta, L. Tom ´as, J. Tordsson, and B. Sanz, “An unsupervised approach to online noisy-neighbor detection in cloud data centers,”Expert Systems with Applications, vol. 89, pp. 188–204, 2017

    work page 2017

  7. [7]

    Mind the gap: Broken promises of cpu reservations in containerized multi-tenant clouds,

    L. Liu, H. Wang, A. Wang, M. Xiao, Y . Cheng, and S. Chen, “Mind the gap: Broken promises of cpu reservations in containerized multi-tenant clouds,” inProceedings of the ACM Symposium on Cloud Computing, 2021, pp. 243–257

    work page 2021

  8. [8]

    Detection of quality of service degradation on multi-tenant container- ized services,

    P. Horchulhack, E. K. Viegas, A. O. Santin, F. V . Ramos, and P. Tedeschi, “Detection of quality of service degradation on multi-tenant container- ized services,”Journal of Network and Computer Applications, vol. 224, p. 103839, 2024

    work page 2024

  9. [9]

    A causality mining and knowledge graph based method of root cause diagnosis for performance anomaly in cloud applications,

    J. Qiu, Q. Du, K. Yin, S.-L. Zhang, and C. Qian, “A causality mining and knowledge graph based method of root cause diagnosis for performance anomaly in cloud applications,”Applied Sciences, vol. 10, no. 6, p. 2166, 2020

    work page 2020

  10. [10]

    Root cause analysis of failures in microservices through causal discovery,

    A. Ikram, S. Chakraborty, S. Mitra, S. Saini, S. Bagchi, and M. Ko- caoglu, “Root cause analysis of failures in microservices through causal discovery,”Advances in Neural Information Processing Systems, vol. 35, pp. 31 158–31 170, 2022

    work page 2022

  11. [11]

    Root cause analysis for microservice system based on causal inference: How far are we?

    L. Pham, H. Ha, and H. Zhang, “Root cause analysis for microservice system based on causal inference: How far are we?” inProceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering, 2024, pp. 706–715

    work page 2024

  12. [12]

    5gc-analyser: Demistifying the 5g core network through statistical analysis,

    A. Khichane, I. Fajjari, N. Aitsaadi, and M. Gueroui, “5gc-analyser: Demistifying the 5g core network through statistical analysis,” inICC 2024-IEEE International Conference on Communications. IEEE, 2024, pp. 3682–3688

    work page 2024

  13. [13]

    Metrics and techniques for quan- tifying performance isolation in cloud environments,

    R. Krebs, C. Momm, and S. Kounev, “Metrics and techniques for quan- tifying performance isolation in cloud environments,” inProceedings of the 8th international acm sigsoft conference on quality of software architectures, 2012, pp. 91–100

    work page 2012

  14. [14]

    A performance evaluation of containers running on managed kubernetes services,

    A. P. Ferreira and R. Sinnott, “A performance evaluation of containers running on managed kubernetes services,” in2019 IEEE international conference on cloud computing technology and science (CloudCom). IEEE, 2019, pp. 199–208

    work page 2019

  15. [15]

    Containers and virtual machines at scale: A comparative study,

    P. Sharma, L. Chaufournier, P. Shenoy, and Y . Tay, “Containers and virtual machines at scale: A comparative study,” inProceedings of the 17th international middleware conference, 2016, pp. 1–13

    work page 2016

  16. [16]

    Zero-touch ai-driven distributed man- agement for energy-efficient 6g massive network slicing,

    H. Chergui, L. Blanco, L. A. Garrido, K. Ramantas, S. Kukli ´nski, A. Ksentini, and C. Verikoukis, “Zero-touch ai-driven distributed man- agement for energy-efficient 6g massive network slicing,”Ieee Network, vol. 35, no. 6, pp. 43–49, 2022

    work page 2022