pith. sign in

arxiv: 2605.15866 · v1 · pith:JFZYJZM6new · submitted 2026-05-15 · 💻 cs.DC

Evaluating Container Orchestration for Neuromorphic Workloads in Virtual Edge Environments

Huyen Pham , Bilhanan Silverajan This is my paper

Pith reviewed 2026-05-19 19:04 UTC · model grok-4.3

classification 💻 cs.DC
keywords spiking neural networkscontainer orchestrationKubernetesedge computingneuromorphic computingresource allocationlatencythroughput
0
0 comments X p. Extension
pith:JFZYJZM6 Add to your LaTeX paper What is a Pith Number? 
\usepackage{pith}
\pithnumber{JFZYJZM6}

Prints a linked pith:JFZYJZM6 badge after your title and writes the identifier into PDF metadata. Compiles on arXiv with no extra files. Learn more

The pith

SNN workloads are highly sensitive to CPU restrictions in Kubernetes-based virtual edge environments, with 0.5 cores causing 47.6 times higher latency.

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

The paper investigates the deployment of spiking neural network workloads using container orchestration in simulated edge settings with Kubernetes. It establishes that these workloads suffer major performance hits from limited CPU, such as a 47.6-fold latency increase and 49-fold throughput drop at 0.5 cores, although accuracy does not change. This is important for bringing energy-efficient neuromorphic computing to resource-constrained edge devices. The results also note that standard orchestration features like round-robin routing can add unwanted delays when scaling multiple instances of the workload.

Core claim

Spiking neural network workloads prove highly sensitive to resource availability when deployed in a single-node K3d Kubernetes cluster on a virtualized Windows host. Limiting CPU to 0.5 cores raises median latency by 47.6 times and cuts throughput by 49 times, with accuracy holding steady in all successful runs. The cluster handles deployment and scaling of the workloads, yet the default routing policy produces high tail latency for scaled replicas, pointing to a mismatch for inference tasks.

What carries the argument

Resource restriction experiments on SNN inference in K3d, measuring latency, throughput, accuracy, and overhead under different CPU and memory allocations.

If this is right

  • Adequate CPU provisioning is necessary to prevent extreme latency increases and throughput losses in containerized SNN deployments.
  • Default Kubernetes load balancing strategies lead to significant tail latencies for concurrent SNN inference replicas.
  • Neuromorphic workloads can be orchestrated successfully if memory limits are not exceeded.
  • SNN classification accuracy remains consistent regardless of resource constraints that permit execution.

Where Pith is reading between the lines

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

  • The findings could extend to physical edge hardware, implying similar resource tuning needs for real deployments.
  • Custom orchestration policies might better match the event-driven nature of SNNs to reduce tail latencies.
  • Future optimizations could focus on memory efficiency to allow more constrained configurations without failure.

Load-bearing premise

A single-node K3d cluster on Windows 11 with WSL2 and Docker Desktop represents typical virtual edge environments and SNN behavior.

What would settle it

Repeating the CPU restriction tests to 0.5 cores on physical edge devices and checking if latency multiplies by approximately 47.6 times and throughput drops by 49 times.

Figures

Figures reproduced from arXiv: 2605.15866 by Bilhanan Silverajan, Huyen Pham.

Figure 1
Figure 1. Figure 1: System overview illustrating training, containerization, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Network architecture of the Diehl and Cook SNN [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Infrastructure overhead (p50 minus empirical floor) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Latency distribution under resource constraints (base [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Throughput under resource constraints. Per-digit mean latency is visualized in [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Mean latency per MNIST digit class across configura [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Latency distribution across all three K3d scenarios. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Latency summary and throughput across all three K3d [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
read the original abstract

The growing adoption of edge computing has created an increasing need for workloads capable of operating under strict resource and energy constraints. Neuromorphic computing, and spiking neural networks (SNNs) in particular, offers an energy-efficient alternative to conventional machine learning through event-driven computation. However, how SNN workloads behave when deployed within modern container orchestration frameworks, especially in edge environments, remains largely unexplored. This paper investigates the feasibility of deploying and orchestrating SNN workloads in a virtual edge environment using Kubernetes, focusing on end-to-end latency, throughput, classification accuracy, infrastructure overhead, and runtime behavior under concurrent load. Experiments were conducted on a single-node K3d cluster running on a Windows 11 host with WSL2 and Docker Desktop. The results show that SNN workloads are highly sensitive to resource availability. Restricting CPU to 0.5 cores increased median latency by 47.6x and reduced throughput by 49x, while the most constrained configuration failed due to insufficient memory. Classification accuracy remained stable across all working configurations. From an orchestration perspective, K3d successfully deployed and scaled SNN workloads, though its default round-robin routing policy introduced significant tail latency under replica scaling, highlighting a mismatch between stateless load-balancing assumptions and long-running inference workloads. Overall, this study provides a baseline for deploying neuromorphic workloads in containerized edge environments and highlights the importance of resource provisioning and orchestration configuration. Future work should explore improved routing strategies, memory optimization, and validation on physical edge hardware.

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

1 major / 2 minor

Summary. The manuscript evaluates the deployment of spiking neural network (SNN) workloads using the K3d Kubernetes distribution in a virtual edge environment simulated via a single-node cluster on Windows 11 with WSL2 and Docker Desktop. It reports that SNN workloads are highly sensitive to CPU resource restrictions, with a 0.5-core limit causing a 47.6x increase in median latency and 49x reduction in throughput while classification accuracy remains stable across viable configurations. The work also notes that K3d can deploy and scale the workloads but that its default round-robin policy produces high tail latency, and positions the study as a baseline for container-orchestrated neuromorphic edge computing.

Significance. If the quantitative sensitivity results hold, the paper supplies a useful empirical baseline on resource provisioning for event-driven neuromorphic workloads under container orchestration, an area that has received limited prior attention. The identification of a mismatch between stateless load-balancing assumptions and long-running inference tasks is a practical observation that could inform future orchestration designs for edge AI.

major comments (1)
  1. Abstract and experimental setup description: The central claims of 47.6x median latency increase and 49x throughput drop at 0.5 CPU cores rest entirely on measurements from a single-node K3d cluster atop Windows 11 + WSL2 + Docker Desktop. This stack introduces additional scheduling, memory-mapping, and I/O layers absent from native Linux/ARM edge nodes. No direct comparison (e.g., cgroups-only vs. full virtualization or perf-counter isolation of host overhead) is reported, so the observed multipliers may partly reflect the experimental platform rather than intrinsic SNN behavior under container orchestration. This directly affects the interpretation of the results as a model for virtual edge environments.
minor comments (2)
  1. Abstract: The statement that 'classification accuracy remained stable across all working configurations' would be strengthened by reporting the number of trials, variance, or confidence intervals; the current phrasing leaves the robustness of the stability claim unclear.
  2. The manuscript mentions 'infrastructure overhead' as one of the investigated aspects but does not supply concrete metrics (e.g., CPU or memory overhead percentages) for this quantity in the results summary.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for acknowledging the potential value of this work as an empirical baseline. We address the major comment below.

read point-by-point responses

  1. Referee: Abstract and experimental setup description: The central claims of 47.6x median latency increase and 49x throughput drop at 0.5 CPU cores rest entirely on measurements from a single-node K3d cluster atop Windows 11 + WSL2 + Docker Desktop. This stack introduces additional scheduling, memory-mapping, and I/O layers absent from native Linux/ARM edge nodes. No direct comparison (e.g., cgroups-only vs. full virtualization or perf-counter isolation of host overhead) is reported, so the observed multipliers may partly reflect the experimental platform rather than intrinsic SNN behavior under container orchestration. This directly affects the interpretation of the results as a model for virtual edge environments.

    Authors: We agree that the chosen platform (single-node K3d on Windows 11 + WSL2 + Docker Desktop) introduces virtualization layers not present on native Linux/ARM edge hardware, and that no isolation experiments (e.g., cgroups-only or perf-counter measurements) were performed to quantify host overhead separately. The manuscript is explicitly scoped to virtual edge environments, and this stack was selected to enable reproducible container-orchestration experiments with K3d. The reported sensitivity results therefore characterize SNN behavior under the full containerized virtual setup rather than claiming to isolate intrinsic neuromorphic properties. In revision we will add an explicit limitations paragraph in the experimental setup section that discusses these platform-specific factors and their possible contribution to the observed multipliers. We will also expand the future-work paragraph to prioritize native-hardware validation. These textual changes will clarify the scope without altering the core claims. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical benchmarking with direct measurements

full rationale

The paper reports experimental results from running SNN workloads on a single-node K3d cluster, measuring latency, throughput, accuracy, and orchestration behavior under CPU/memory restrictions. No derivations, equations, predictive models, or first-principles claims appear in the abstract or described content. All reported multipliers (e.g., 47.6x latency increase) are direct observations from the testbed rather than outputs of any fitted function or self-referential construction. The study is therefore self-contained against external benchmarks with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on experimental observations in a specific virtual setup with no new theoretical entities or fitted parameters beyond the test configurations.

axioms (1)
  • domain assumption The virtual K3d environment on WSL2 accurately simulates edge computing conditions for SNN workloads
    The generalization of results to real edge hardware depends on this assumption.

pith-pipeline@v0.9.0 · 5803 in / 1360 out tokens · 70433 ms · 2026-05-19T19:04:35.069925+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 2 internal anchors

  1. [1]

    Neuro-inspired task offloading in edge-IoT networks using spiking neural networks

    F. D. Rossi, “Neuro-inspired task offloading in edge-IoT networks using spiking neural networks.” [Online]. Available: http://arxiv.org/ abs/2511.01127

    work page arXiv

  2. [2]

    Edge intelligence with spiking neural networks

    S. Deng, D. Yu, C. Lv, X. Du, L. Jiang, X. Zhao, W. Tong, X. Zheng, W. Fang, P. Zhao, G. Pan, S. Dustdar, and A. Y . Zomaya, “Edge intelligence with spiking neural networks.” [Online]. Available: http://arxiv.org/abs/2507.14069

    work page arXiv

  3. [3]

    Container orchestration and kubernetes enhancements,

    S. Das, “Container orchestration and kubernetes enhancements,” vol. 16, no. 1. [Online]. Available: https://www.ijsat.org/research-paper.php?id= 2594

    work page

  4. [4]

    jennyhpham/distributed neuro sim,

    J. Pham, “jennyhpham/distributed neuro sim,” original-date: 2025- 12-10T21:06:43Z. [Online]. Available: https://github.com/jennyhpham/ distributed neuro sim

    work page 2025

  5. [5]

    Neuromorphic computing at scale,

    D. Kudithipudi, C. Schuman, C. M. Vineyard, T. Pandit, C. Merkel, R. Kubendran, J. B. Aimone, G. Orchard, C. Mayr, R. Benosman, J. Hays, C. Young, C. Bartolozzi, A. Majumdar, S. G. Cardwell, M. Payvand, S. Buckley, S. Kulkarni, H. A. Gonzalez, G. Cauwenberghs, C. S. Thakur, A. Subramoney, and S. Furber, “Neuromorphic computing at scale,” vol. 637, no. 804...

    work page

  6. [6]

    A survey on neuromorphic computing: Models and hardware,

    A. Shrestha, H. Fang, Z. Mei, D. P. Rider, Q. Wu, and Q. Qiu, “A survey on neuromorphic computing: Models and hardware,” vol. 22, no. 2, pp. 6–35. [Online]. Available: https: //ieeexplore.ieee.org/abstract/document/9782767

    work page arXiv

  7. [7]

    Exploring neuromorphic computing based on spiking neural networks: Algorithms to hardware,

    N. Rathi, I. Chakraborty, A. Kosta, A. Sengupta, A. Ankit, P. Panda, and K. Roy, “Exploring neuromorphic computing based on spiking neural networks: Algorithms to hardware,” vol. 55, no. 12, pp. 243:1–243:49. [Online]. Available: https://dl.acm.org/doi/10.1145/3571155

    work page doi:10.1145/3571155

  8. [8]

    Exploring temporal information dynamics in spiking neural networks

    Y . Kim, Y . Li, H. Park, Y . Venkatesha, A. Hambitzer, and P. Panda, “Exploring temporal information dynamics in spiking neural networks.” [Online]. Available: http://arxiv.org/abs/2211.14406

    work page arXiv

  9. [9]

    Spiking Deep Networks with LIF Neurons

    E. Hunsberger and C. Eliasmith, “Spiking deep networks with LIF neurons.” [Online]. Available: http://arxiv.org/abs/1510.08829

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Spiking neural networks and their applications: A review,

    K. Yamazaki, V .-K. V o-Ho, D. Bulsara, and N. Le, “Spiking neural networks and their applications: A review,” vol. 12, no. 7, p. 863, publisher: Multidisciplinary Digital Publishing Institute. [Online]. Available: https://www.mdpi.com/2076-3425/12/7/863

    work page 2076

  11. [11]

    Learning rules in spiking neural networks: A survey,

    Z. Yi, J. Lian, Q. Liu, H. Zhu, D. Liang, and J. Liu, “Learning rules in spiking neural networks: A survey,” vol. 531, pp. 163–179. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0925231223001662

    work page

  12. [12]

    An unsupervised STDP-based spiking neural network inspired by biologically plausible learning rules and connections,

    Y . Dong, D. Zhao, Y . Li, and Y . Zeng, “An unsupervised STDP-based spiking neural network inspired by biologically plausible learning rules and connections,” vol. 165, pp. 799–808. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0893608023003301

    work page

  13. [13]

    TrueNorth: Accelerating from zero to 64 million neurons in 10 years,

    M. V . DeBole, B. Taba, A. Amir, F. Akopyan, A. Andreopoulos, W. P. Risk, J. Kusnitz, C. Ortega Otero, T. K. Nayak, R. Appuswamy, P. J. Carlson, A. S. Cassidy, P. Datta, S. K. Esser, G. J. Garreau, K. L. Holland, S. Lekuch, M. Mastro, J. McKinstry, C. di Nolfo, B. Paulovicks, J. Sawada, K. Schleupen, B. G. Shaw, J. L. Klamo, M. D. Flickner, J. V . Arthur,...

    work page arXiv

  14. [14]

    Loihi: A neuromorphic manycore processor with on-chip learning,

    M. Davies, N. Srinivasa, T.-H. Lin, G. Chinya, Y . Cao, S. H. Choday, G. Dimou, P. Joshi, N. Imam, S. Jain, Y . Liao, C.-K. Lin, A. Lines, R. Liu, D. Mathaikutty, S. McCoy, A. Paul, J. Tse, G. Venkataramanan, Y .-H. Weng, A. Wild, Y . Yang, and H. Wang, “Loihi: A neuromorphic manycore processor with on-chip learning,” vol. 38, no. 1, pp. 82–99. [Online]. ...

    work page arXiv

  15. [15]

    SpiNNaker 2: A 10 million core processor system for brain simulation and machine learning

    C. Mayr, S. Hoeppner, and S. Furber, “SpiNNaker 2: A 10 million core processor system for brain simulation and machine learning.” [Online]. Available: http://arxiv.org/abs/1911.02385

    work page arXiv 1911

  16. [16]

    BindsNET: A machine learning-oriented spiking neural networks library in python,

    H. Hazan, D. J. Saunders, H. Khan, D. Patel, D. T. Sanghavi, H. T. Siegelmann, and R. Kozma, “BindsNET: A machine learning-oriented spiking neural networks library in python,” vol. 12. [Online]. Available: https://www.frontiersin.org/journals/ neuroinformatics/articles/10.3389/fninf.2018.00089/full

    work page doi:10.3389/fninf.2018.00089/full 2018

  17. [17]

    Edge computing: Vision and challenges,

    W. Shi, J. Cao, Q. Zhang, Y . Li, and L. Xu, “Edge computing: Vision and challenges,” vol. 3, no. 5, pp. 637–646. [Online]. Available: https://ieeexplore.ieee.org/abstract/document/7488250

    work page arXiv

  18. [18]

    An overview on edge computing research,

    K. Cao, Y . Liu, G. Meng, and Q. Sun, “An overview on edge computing research,” vol. 8, pp. 85 714–85 728. [Online]. Available: https://ieeexplore.ieee.org/document/9083958

    work page arXiv

  19. [19]

    A comprehensive review on edge computing - architecture, algorithm, and devices,

    A. G. P, A. Madhumitha, B. P. Swathi, G. R. M, and B. Vanaja, “A comprehensive review on edge computing - architecture, algorithm, and devices,” vol. 11, no. 11, pp. 545–553. [Online]. Available: https://ijirt.org/IJIRT|AnUGCCompliantPeerreviewedJournal| InternationalOpenAccessJournal

    work page

  20. [20]

    A survey on open-source edge computing simulators and emulators: The computing and networking convergence perspective

    J. Qi, C. Liu, X. Zhang, L. Wang, R. Wang, J. Dong, and Y . Yu, “A survey on open-source edge computing simulators and emulators: The computing and networking convergence perspective.” [Online]. Available: http://arxiv.org/abs/2505.09995

    work page arXiv

  21. [21]

    Fogify: A fog computing emulation framework,

    M. Symeonides, Z. Georgiou, D. Trihinas, G. Pallis, and M. D. Dikaiakos, “Fogify: A fog computing emulation framework,” in2020 IEEE/ACM Symposium on Edge Computing (SEC), pp. 42–54. [Online]. Available: https://ieeexplore.ieee.org/document/9355701

    work page arXiv

  22. [22]

    Borg, omega, and kubernetes,

    B. Burns, B. Grant, D. Oppenheimer, E. Brewer, and J. Wilkes, “Borg, omega, and kubernetes,” vol. 59, no. 5, pp. 50–57. [Online]. Available: https://dl.acm.org/doi/10.1145/2890784

    work page doi:10.1145/2890784

  23. [23]

    Skydiver: A spiking neural network accelerator exploiting spatio-temporal workload balance,

    Q. Chen, C. Gao, X. Fang, and H. Luan, “Skydiver: A spiking neural network accelerator exploiting spatio-temporal workload balance,” vol. 41, no. 12, pp. 5732–5736. [Online]. Available: http://arxiv.org/ abs/2203.07516

    work page arXiv

  24. [24]

    EdgeMap: An optimized mapping toolchain for spiking neural network in edge computing,

    J. Xue, L. Xie, F. Chen, L. Wu, Q. Tian, Y . Zhou, R. Ying, and P. Liu, “EdgeMap: An optimized mapping toolchain for spiking neural network in edge computing,” vol. 23, no. 14. [Online]. Available: https://www.mdpi.com/1424-8220/23/14/6548

    work page

  25. [25]

    Performance evaluation of container orchestration tools in edge computing environments,

    I. ˇCili´c, P. Krivi ´c, I. Podnar ˇZarko, and M. Ku ˇsek, “Performance evaluation of container orchestration tools in edge computing environments,” vol. 23, no. 8, p. 4008, publisher: Multidisciplinary Digital Publishing Institute. [Online]. Available: https://www.mdpi.com/ 1424-8220/23/8/4008

    work page

  26. [26]

    Performance Evaluation of Deep Learning Tools in Docker Containers

    P. Xu, S. Shi, and X. Chu, “Performance evaluation of deep learning tools in docker containers.” [Online]. Available: http: //arxiv.org/abs/1711.03386

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    A scalable NorthPole system with end-to-end vertical integration for low-latency and energy-efficient LLM inference,

    M. V . DeBole, R. Appuswamy, N. McGlohon, B. Taba, S. K. Esser, F. Akopyan, J. V . Arthur, A. Amir, A. Andreopoulos, P. J. Carlson, A. S. Cassidy, P. Datta, M. D. Flickner, R. Gandhasri, G. J. Garreau, M. Ito, J. L. Klamo, J. A. Kusnitz, N. J. McClatchey, J. L. McKinstry, T. K. Nayak, C. O. Otero, H. Penner, W. P. Risk, J. Sawada, J. Sivagnaname, D. F. Sm...

    work page arXiv

  28. [28]

    [Online]

    MNIST dataset. [Online]. Available: https://www.kaggle.com/datasets/ hojjatk/mnist-dataset

    work page

  29. [29]

    Unsupervised learning of digit recognition using spike-timing-dependent plasticity,

    P. U. Diehl and M. Cook, “Unsupervised learning of digit recognition using spike-timing-dependent plasticity,” vol. 9. [Online]. Avail- able: https://www.frontiersin.org/journals/computational-neuroscience/ articles/10.3389/fncom.2015.00099/full

    work page doi:10.3389/fncom.2015.00099/full 2015