LearnedCache: An eBPF-Integrated Perceptron-Based Eviction Policy for the Linux Page Cache

Zejia Qi

arxiv: 2605.26168 · v1 · pith:UCXHB7PXnew · submitted 2026-05-25 · 💻 cs.OS · cs.LG

LearnedCache: An eBPF-Integrated Perceptron-Based Eviction Policy for the Linux Page Cache

Zejia Qi This is my paper

Pith reviewed 2026-06-29 20:00 UTC · model grok-4.3

classification 💻 cs.OS cs.LG

keywords Linux page cacheeBPFperceptroncache evictionmachine learninginsertion ratepage reuse timeFIFO

0 comments

The pith

A perceptron model embedded via eBPF outperforms FIFO eviction by up to 10 percent insertion rate in Linux page cache under tested workloads.

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

The paper builds LearnedCache by training single-layer perceptrons on real kernel traces to predict page reuse times, then loads the resulting models into the running Linux kernel through eBPF so they can drive eviction decisions in real time. Experiments across multiple workloads compare the approach against FIFO using 50 paired trials per workload and report median AUC near 80 percent for the reuse-time models together with statistically significant gains in insertion rate of up to 10 percent in some cases. The work shows that these machine-learning policies can be executed inside the kernel with only minimal overhead. A reader would care because the page cache sits on the critical path for every disk access, so replacing fixed heuristics with workload-specific predictors could reduce I/O without changing application code.

Core claim

LearnedCache trains linear perceptron models offline on kernel-collected page traces from diverse workloads, embeds the compact models inside the kernel via eBPF, and uses them at eviction time to rank pages by predicted reuse distance; statistical tests against FIFO show the ML policy produces higher insertion rates (a frequency-adjusted hit-rate metric) by margins reaching 10 percent in specific workloads while adding negligible runtime cost.

What carries the argument

The single-layer perceptron that models page reuse time and is executed inside the kernel by an eBPF program attached to the page-cache eviction path.

If this is right

Machine-learning eviction policies become feasible inside the production Linux kernel rather than only in user-space applications.
Insertion-rate gains of several percent translate directly into fewer disk reads under the tested access patterns.
Statistical significance testing over repeated paired trials provides evidence that the improvements are repeatable rather than noise.
The same offline-training-plus-eBPF-embedding pattern can be applied to other kernel decisions that currently rely on fixed heuristics.

Where Pith is reading between the lines

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

If the same embedding technique works for perceptrons, it could also support slightly larger models or online retraining inside the kernel.
Workloads where the perceptron underperforms FIFO might benefit from ensemble or workload-specific model selection at load time.
The approach opens a route for other operating systems to adopt kernel-resident ML policies without rewriting core data structures.

Load-bearing premise

Models trained offline on traces from several workloads can be loaded and executed in real time through eBPF without introducing kernel bugs or overhead that erase the measured performance gains.

What would settle it

A kernel run in which the eBPF program either panics, consumes more than a few percent extra CPU, or produces insertion rates statistically indistinguishable from or worse than FIFO across the same workload set.

Figures

Figures reproduced from arXiv: 2605.26168 by Zejia Qi.

**Figure 1.** Figure 1: Tracer and evaluation harness workflow. pairwise linear ranker implemented as a perceptron. This model architecture was chosen to maximize performance as it boils down to a vector dot product between the feature vector and the weight vector, minimizing the number of operations when being run multiple times in the kernel. The entire model was built using scikit-learn, Keras, and TensorFlow. 1) The Discretiz… view at source ↗

**Figure 2.** Figure 2: Skewed feature distributions motivating discretization. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 4.** Figure 4: Learned weights for varmail (green=keep, red=evict) [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Learned weights for webserver (green=keep, red=evict) [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 3.** Figure 3: AUC and F1 score by workload [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 7.** Figure 7: displays the model loading flow. Each component will be described [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗

**Figure 8.** Figure 8: Webproxy evaluation plots. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗

**Figure 9.** Figure 9: Webserver evaluation plots. 3) Copyfiles: Copyfiles is a workload that repeatedly copies files from one directory to another. A statistically significant improvement is observed. The p-value of 1.68×10−2 is less than α = 0.05, so we can reject the null hypothesis. A raw insertion rate decrease of 0.223 is found, with an effect size of −0.350 [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗

**Figure 11.** Figure 11: Varmail evaluation plots. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗

**Figure 10.** Figure 10: Copyfiles evaluation plots. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗

**Figure 12.** Figure 12: Openfiles evaluation plots. to zero, and parity lies well within the constructed 95% confidence interval. 6) Mongo: Mongo is a small workload designed to simulate small chunked accesses of a MongoDB-like database. No statistically significant regression is observed. The pvalue of 5.99 × 10−2 is greater than α = 0.05, so we fail to reject the null hypothesis. A raw insertion rate increase of 0.225 is fou… view at source ↗

**Figure 14.** Figure 14: Eviction decision latency comparison. This difference is statistically significant. RAM access is typically hundreds of thousands of times faster than reading from disk. Assuming the underlying machine matches these estimates and does not produce massive amounts of page cache churn, the milliseconds of time saved would outweigh the model overhead. VII. Conclusions LearnedCache demonstrates that performant… view at source ↗

**Figure 13.** Figure 13: Mongo evaluation plots. As shown in [PITH_FULL_IMAGE:figures/full_fig_p010_13.png] view at source ↗

read the original abstract

Linux is the foundation of the digital age, accounting for the majority of the cloud and mobile OS markets. Any device that runs Linux uses the Linux page cache, a central pillar in OS and application performance, serving to reduce extraneous disk access. Many page cache eviction policies have been developed but remain bound by the rigidity of heuristics. The rise of AI-driven tools in recent years, melded with the ever-increasing variety of workloads for Linux devices, sets the stage for machine-learning-driven cache eviction policies. Promising research has been done in this field, but only in the field of user-space applications such as CDNs. We develop LearnedCache, an eBPF-integrated single-layer perceptron-based cache eviction policy for the Linux page cache, trained on real kernel data from diverse workloads. We demonstrate median AUCs of nearly 80% over multiple linear models modeling page reuse time, then take a step further by embedding these models inside the Linux kernel for real-time performance evaluation. Through statistical testing over 50 paired trials against a baseline of FIFO for each workload, LearnedCache reveals that machine-learning-derived cache eviction policies are practical in the Linux kernel under representative empirical workloads and are able to surpass conventional FIFO by statistically significant margins of up to 10% in insertion rate, a frequency-adjusted derivation of cache hit rate, in specific workloads while incurring minimal overhead.

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.

Desk Editor's Note private letter to a colleague

The abstract describes moving a perceptron-based eviction policy into the Linux page cache via eBPF and reports gains over FIFO, but supplies almost no implementation or overhead data to back the practicality claim.

read the letter

The paper's main step is taking offline-trained single-layer perceptrons for page reuse prediction and embedding them in the kernel eviction path with eBPF. It reports median AUC near 80% and statistically significant insertion-rate improvements up to 10% versus FIFO across 50 paired trials on various workloads, while asserting minimal overhead.

This is new relative to the user-space ML cache work it cites. The evaluation framing as direct empirical comparison against a simple baseline is straightforward.

The soft spots are substantial and sit right at the central claim. The abstract gives no account of training data collection, model selection, exclusion criteria, or the precise statistical tests. It also provides zero kernel-path latency figures, no description of how the eBPF program is invoked per candidate or per fault, and no discussion of verifier limits or map costs. Without those, the "minimal overhead" and "practical" conclusions cannot be checked.

The stress-test concern about unverified eBPF overhead and integration correctness holds up on the supplied text. If per-decision cost is more than a few hundred nanoseconds or if the policy cannot safely coexist with existing structures, the reported gains do not demonstrate practicality.

This is for people tracking kernel-ML intersections or cache policy experiments. It deserves peer review so the implementation details and raw measurements can be examined; the idea is worth testing even if the current write-up is too thin to stand on its own.

Referee Report

2 major / 1 minor

Summary. The manuscript presents LearnedCache, an eBPF-integrated single-layer perceptron model for Linux page cache eviction. Models are trained offline on real kernel data from diverse workloads to predict page reuse time (median AUC near 80%). The work embeds these models in the kernel and evaluates them via 50 paired trials per workload against a FIFO baseline, reporting statistically significant gains of up to 10% in insertion rate (a frequency-adjusted cache hit rate metric) in specific workloads while claiming minimal overhead.

Significance. If the embedding, overhead, and statistical claims hold after verification, the result would demonstrate that simple ML models can be practically integrated into the Linux kernel's page cache path and outperform conventional heuristics under representative workloads. Strengths include the use of real kernel-collected training data, direct eBPF embedding for real-time evaluation, and the paired-trial statistical design. This could support broader exploration of learned policies in OS components.

major comments (2)

[Abstract] Abstract: the central practicality claim requires that the perceptron executes in the real-time eviction path without overhead that erodes the reported insertion-rate gains, yet no per-decision latency figures, invocation details (per eviction candidate or page fault), eBPF map access costs, or verifier-limit discussion are supplied. This directly undermines the 'minimal overhead' assertion and the conclusion that the AUC results translate to net kernel benefit.
[Abstract] Abstract: the reported median AUC near 80%, 10% insertion-rate gains, and statistical significance over 50 paired trials rest on unspecified training procedures, data-collection methods, page-exclusion rules, and exact statistical tests. Without these, the soundness of the empirical comparison to FIFO cannot be verified and the main claim remains uncheckable.

minor comments (1)

[Abstract] Abstract: 'insertion rate' is described as a frequency-adjusted derivation of cache hit rate, but an explicit equation or definition would aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for these focused comments on the abstract. Both points identify areas where additional detail would strengthen verifiability and the practicality claim. We will revise the manuscript to supply the requested information.

read point-by-point responses

Referee: [Abstract] Abstract: the central practicality claim requires that the perceptron executes in the real-time eviction path without overhead that erodes the reported insertion-rate gains, yet no per-decision latency figures, invocation details (per eviction candidate or page fault), eBPF map access costs, or verifier-limit discussion are supplied. This directly undermines the 'minimal overhead' assertion and the conclusion that the AUC results translate to net kernel benefit.

Authors: We agree that the abstract does not contain per-decision latency numbers, invocation frequency, map-access costs, or verifier-limit discussion. The current manuscript reports aggregate overhead but lacks these granular figures. We will add measured per-decision latencies (in cycles and nanoseconds), clarify that the model is invoked once per eviction candidate, quantify eBPF map lookup costs, and discuss verifier constraints in a revised abstract and a new methods paragraph. This revision will directly support the minimal-overhead claim. revision: yes
Referee: [Abstract] Abstract: the reported median AUC near 80%, 10% insertion-rate gains, and statistical significance over 50 paired trials rest on unspecified training procedures, data-collection methods, page-exclusion rules, and exact statistical tests. Without these, the soundness of the empirical comparison to FIFO cannot be verified and the main claim remains uncheckable.

Authors: We accept that the abstract omits these methodological specifics, rendering the numerical claims difficult to verify from the abstract alone. We will expand the abstract with concise statements on training (offline supervised learning on kernel-collected reuse traces), data-collection procedure, page-exclusion criteria, and the statistical test (paired Wilcoxon signed-rank test across 50 trials). A corresponding methods subsection will provide full procedural detail so that the AUC, insertion-rate gains, and significance claims become reproducible. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical evaluation is self-contained

full rationale

The paper trains single-layer perceptrons offline on kernel-collected reuse-time data from diverse workloads, embeds the models via eBPF, and reports median AUC ~80% plus insertion-rate gains versus FIFO measured over 50 paired trials per workload. No equations, self-citations, or derivations are present that reduce the reported performance margins to fitted parameters by construction or to prior author work. The central claim is therefore an external empirical result rather than a tautology.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no free parameters, axioms, or invented entities are described in sufficient detail to populate the ledger.

pith-pipeline@v0.9.1-grok · 5770 in / 1015 out tokens · 33988 ms · 2026-06-29T20:00:09.706152+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 9 canonical work pages · 2 internal anchors

[1]

Linux cloud infrastructure (CSE-41369),

UC San Diego Division of Extended Studies, “Linux cloud infrastructure (CSE-41369),” UC San Diego Division of Extended Studies, accessed: 2026-02-20. [Online]. Available: https://extendedstudies.ucsd.edu/courses/ linux-cloud-infrastructure-cse-41369

2026
[2]

Android global market share statistics,

CommandLinux, “Android global market share statistics,” CommandLinux, accessed: 2026-02-20. [Online]. Available: https://commandlinux.com/android/ android-global-market-share-statistics/

2026
[3]

Operating system support for database management,

M. Stonebraker, “Operating system support for database management,”Communications of the ACM, vol. 24, no. 7, pp. 412–418, Jul. 1981. [Online]. Available: https://doi.org/10.1145/ 358699.358703

work page arXiv 1981
[4]

The multi-generational LRU,

J. Corbet, “The multi-generational LRU,” LWN.net, Apr. 2021, accessed: 2026-02-20. [Online]. Available: https://lwn.net/ Articles/851184/

2021
[5]

P2Cache: An application-directed page cache for improving performance of data-intensive applications,

D. Lee, I. Choi, C. Lee, S. Lee, and J. Kim, “P2Cache: An application-directed page cache for improving performance of data-intensive applications,” inProceedings of the 15th ACM Workshop on Hot Topics in Storage and File Systems (HotStorage ’23). New York, NY, USA: Association for Computing Machinery, Jul. 2023, pp. 31–36. [Online]. Available: https://do...

work page doi:10.1145/3599691.3603408 2023
[6]

vmlinux.h

T. Zussman, I. Zarkadas, J. Carin, A. Cheng, H. Franke, J. Pfefferle, and A. Cidon, “cache_ext: Customizing the page cache with eBPF,” inProceedings of the ACM SIGOPS 31st Symposium on Operating Systems Principles (SOSP ’25), ser. SOSP ’25. New York, NY, USA: Association for Computing Machinery, Oct. 2025, pp. 462–478, published as cache_ext; arXiv prepri...

work page doi:10.1145/3731569.3764820 2025
[7]

The eBPF runtime in the Linux kernel,

B. Gbadamosi, L. Leonardi, T. Pulls, T. Høiland-Jørgensen, S. Ferlin-Reiter, S. Sorce, and A. Brunström, “The eBPF runtime in the Linux kernel,” 2024, arXiv:2410.00026. [Online]. Available: https://arxiv.org/abs/2410.00026

work page arXiv 2024
[8]

LHD: Improving cache hit rate by maximizing hit density,

N. Beckmann, H. Chen, and A. Cidon, “LHD: Improving cache hit rate by maximizing hit density,” in15th USENIX Symposium on Networked Systems Design and Implementation (NSDI 18). Renton, WA: USENIX Association, Apr. 2018, pp. 389–403. [Online]. Available: https://www.usenix.org/conference/nsdi18/ presentation/beckmann

2018
[9]

ML-CLOCK: Efficient page cache algorithm based on perceptron-based neural network,

M. Cho and D. Kang, “ML-CLOCK: Efficient page cache algorithm based on perceptron-based neural network,” Electronics, vol. 10, no. 20, p. 2503, 2021. [Online]. Available: https://www.mdpi.com/2079-9292/10/20/2503

2021
[10]

A learned cache eviction framework with minimal overhead,

D. Yang, D. S. Berger, K. Li, and W. Lloyd, “A learned cache eviction framework with minimal overhead,” 2023, arXiv:2301.11886. [Online]. Available: https://arxiv.org/abs/ 2301.11886

work page arXiv 2023
[11]

HALP: Heuristic aided learned preference eviction policy for YouTube content delivery network,

Z. Song, K. Chen, N. Sarda, D. Altınbüken, E. Brevdo, J. Coleman, X. Ju, P. Jurczyk, R. Schooler, and R. Gummadi, “HALP: Heuristic aided learned preference eviction policy for YouTube content delivery network,” in20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). Boston, MA: USENIX Association, Apr. 2023, pp. 1149–1163. [Onlin...

2023
[12]

eBPF: Introduction, tutorials, and community resources,

eBPF Community, “eBPF: Introduction, tutorials, and community resources,” eBPF.io, 2024, accessed: 2026-03-25. [Online]. Available: https://ebpf.io/

2024
[13]

Silberschatz, P

A. Silberschatz, P. B. Galvin, and G. Gagne,Operating System Concepts, 10th ed. Hoboken, NJ: Wiley, 2018

2018
[14]

The state of the page in 2025,

J. Corbet, “The state of the page in 2025,” LWN.net, Mar. 2025, accessed: 2026-03-25. [Online]. Available: https: //lwn.net/Articles/1015320/

work page arXiv 2025
[15]

BPF design Q&A,

The Linux Kernel Organization, “BPF design Q&A,” Linux kernel documentation, accessed: 2026-03-25. [Online]. Available: https://docs.kernel.org/bpf/bpf_design_QA.html

2026
[16]

eBPF verifier,

——, “eBPF verifier,” Linux kernel documentation, accessed: 2026-03-25. [Online]. Available: https://docs.kernel.org/bpf/ verifier.html

2026
[17]

Floating-point API,

——, “Floating-point API,” Linux kernel documentation, accessed: 2026-03-25. [Online]. Available: https://docs.kernel. org/core-api/floating-point.html

2026
[18]

Filebench: A flexible framework for file system benchmarking,

V. Tarasov, E. Zadok, and S. Shepler, “Filebench: A flexible framework for file system benchmarking,”;login: The USENIX Magazine, vol. 41, no. 1, pp. 6–12, 2016, spring 2016. [Online]. Available: https://www.usenix.org/publications/login/ spring2016/tarasov

2016
[19]

folio_mark_accessed() — linux kernel source,

The Linux Kernel Organization, “ folio_mark_accessed() — linux kernel source,” Linux kernel source code, 2023, linux kernel v6.6.8; Accessed: 2026-03-25. [Online]. Avail- able: https://git.kernel.org/pub/scm/linux/kernel/git/stable/ linux.git/tree/mm/swap.c?h=v6.6.8

2023
[20]

Rank analysis of incomplete block designs: I. the method of paired comparisons,

R. A. Bradley and M. E. Terry, “Rank analysis of incomplete block designs: I. the method of paired comparisons,”Biometrika, vol. 39, no. 3/4, pp. 324–345, 1952. [Online]. Available: https://doi.org/10.1093/biomet/39.3-4.324

work page doi:10.1093/biomet/39.3-4.324 1952
[21]

Lecture 24: The Bradley-Terry model,

Z. Fan, “Lecture 24: The Bradley-Terry model,” STATS 200: Introduction to Statistical Inference, Stanford University, 2016, autumn 2016. [Online]. Available: https://web.stanford.edu/ class/archive/stats/stats200/stats200.1172/Lecture24.pdf

2016
[22]

Adam: A Method for Stochastic Optimization

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inInternational Conference on Learning Representations (ICLR), 2015, arXiv:1412.6980. [Online]. Available: https://arxiv.org/abs/1412.6980

work page internal anchor Pith review Pith/arXiv arXiv 2015
[23]

SciPy 1.0--Fundamental Algorithms for Scientific Computing in Python

P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, İ. Polat, Y. Feng, E. W. Moore, J. VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Henriksen, ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1038/s41592-019-0686-2 2020

[1] [1]

Linux cloud infrastructure (CSE-41369),

UC San Diego Division of Extended Studies, “Linux cloud infrastructure (CSE-41369),” UC San Diego Division of Extended Studies, accessed: 2026-02-20. [Online]. Available: https://extendedstudies.ucsd.edu/courses/ linux-cloud-infrastructure-cse-41369

2026

[2] [2]

Android global market share statistics,

CommandLinux, “Android global market share statistics,” CommandLinux, accessed: 2026-02-20. [Online]. Available: https://commandlinux.com/android/ android-global-market-share-statistics/

2026

[3] [3]

Operating system support for database management,

M. Stonebraker, “Operating system support for database management,”Communications of the ACM, vol. 24, no. 7, pp. 412–418, Jul. 1981. [Online]. Available: https://doi.org/10.1145/ 358699.358703

work page arXiv 1981

[4] [4]

The multi-generational LRU,

J. Corbet, “The multi-generational LRU,” LWN.net, Apr. 2021, accessed: 2026-02-20. [Online]. Available: https://lwn.net/ Articles/851184/

2021

[5] [5]

P2Cache: An application-directed page cache for improving performance of data-intensive applications,

D. Lee, I. Choi, C. Lee, S. Lee, and J. Kim, “P2Cache: An application-directed page cache for improving performance of data-intensive applications,” inProceedings of the 15th ACM Workshop on Hot Topics in Storage and File Systems (HotStorage ’23). New York, NY, USA: Association for Computing Machinery, Jul. 2023, pp. 31–36. [Online]. Available: https://do...

work page doi:10.1145/3599691.3603408 2023

[6] [6]

vmlinux.h

T. Zussman, I. Zarkadas, J. Carin, A. Cheng, H. Franke, J. Pfefferle, and A. Cidon, “cache_ext: Customizing the page cache with eBPF,” inProceedings of the ACM SIGOPS 31st Symposium on Operating Systems Principles (SOSP ’25), ser. SOSP ’25. New York, NY, USA: Association for Computing Machinery, Oct. 2025, pp. 462–478, published as cache_ext; arXiv prepri...

work page doi:10.1145/3731569.3764820 2025

[7] [7]

The eBPF runtime in the Linux kernel,

B. Gbadamosi, L. Leonardi, T. Pulls, T. Høiland-Jørgensen, S. Ferlin-Reiter, S. Sorce, and A. Brunström, “The eBPF runtime in the Linux kernel,” 2024, arXiv:2410.00026. [Online]. Available: https://arxiv.org/abs/2410.00026

work page arXiv 2024

[8] [8]

LHD: Improving cache hit rate by maximizing hit density,

N. Beckmann, H. Chen, and A. Cidon, “LHD: Improving cache hit rate by maximizing hit density,” in15th USENIX Symposium on Networked Systems Design and Implementation (NSDI 18). Renton, WA: USENIX Association, Apr. 2018, pp. 389–403. [Online]. Available: https://www.usenix.org/conference/nsdi18/ presentation/beckmann

2018

[9] [9]

ML-CLOCK: Efficient page cache algorithm based on perceptron-based neural network,

M. Cho and D. Kang, “ML-CLOCK: Efficient page cache algorithm based on perceptron-based neural network,” Electronics, vol. 10, no. 20, p. 2503, 2021. [Online]. Available: https://www.mdpi.com/2079-9292/10/20/2503

2021

[10] [10]

A learned cache eviction framework with minimal overhead,

D. Yang, D. S. Berger, K. Li, and W. Lloyd, “A learned cache eviction framework with minimal overhead,” 2023, arXiv:2301.11886. [Online]. Available: https://arxiv.org/abs/ 2301.11886

work page arXiv 2023

[11] [11]

HALP: Heuristic aided learned preference eviction policy for YouTube content delivery network,

Z. Song, K. Chen, N. Sarda, D. Altınbüken, E. Brevdo, J. Coleman, X. Ju, P. Jurczyk, R. Schooler, and R. Gummadi, “HALP: Heuristic aided learned preference eviction policy for YouTube content delivery network,” in20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). Boston, MA: USENIX Association, Apr. 2023, pp. 1149–1163. [Onlin...

2023

[12] [12]

eBPF: Introduction, tutorials, and community resources,

eBPF Community, “eBPF: Introduction, tutorials, and community resources,” eBPF.io, 2024, accessed: 2026-03-25. [Online]. Available: https://ebpf.io/

2024

[13] [13]

Silberschatz, P

A. Silberschatz, P. B. Galvin, and G. Gagne,Operating System Concepts, 10th ed. Hoboken, NJ: Wiley, 2018

2018

[14] [14]

The state of the page in 2025,

J. Corbet, “The state of the page in 2025,” LWN.net, Mar. 2025, accessed: 2026-03-25. [Online]. Available: https: //lwn.net/Articles/1015320/

work page arXiv 2025

[15] [15]

BPF design Q&A,

The Linux Kernel Organization, “BPF design Q&A,” Linux kernel documentation, accessed: 2026-03-25. [Online]. Available: https://docs.kernel.org/bpf/bpf_design_QA.html

2026

[16] [16]

eBPF verifier,

——, “eBPF verifier,” Linux kernel documentation, accessed: 2026-03-25. [Online]. Available: https://docs.kernel.org/bpf/ verifier.html

2026

[17] [17]

Floating-point API,

——, “Floating-point API,” Linux kernel documentation, accessed: 2026-03-25. [Online]. Available: https://docs.kernel. org/core-api/floating-point.html

2026

[18] [18]

Filebench: A flexible framework for file system benchmarking,

V. Tarasov, E. Zadok, and S. Shepler, “Filebench: A flexible framework for file system benchmarking,”;login: The USENIX Magazine, vol. 41, no. 1, pp. 6–12, 2016, spring 2016. [Online]. Available: https://www.usenix.org/publications/login/ spring2016/tarasov

2016

[19] [19]

folio_mark_accessed() — linux kernel source,

The Linux Kernel Organization, “ folio_mark_accessed() — linux kernel source,” Linux kernel source code, 2023, linux kernel v6.6.8; Accessed: 2026-03-25. [Online]. Avail- able: https://git.kernel.org/pub/scm/linux/kernel/git/stable/ linux.git/tree/mm/swap.c?h=v6.6.8

2023

[20] [20]

Rank analysis of incomplete block designs: I. the method of paired comparisons,

R. A. Bradley and M. E. Terry, “Rank analysis of incomplete block designs: I. the method of paired comparisons,”Biometrika, vol. 39, no. 3/4, pp. 324–345, 1952. [Online]. Available: https://doi.org/10.1093/biomet/39.3-4.324

work page doi:10.1093/biomet/39.3-4.324 1952

[21] [21]

Lecture 24: The Bradley-Terry model,

Z. Fan, “Lecture 24: The Bradley-Terry model,” STATS 200: Introduction to Statistical Inference, Stanford University, 2016, autumn 2016. [Online]. Available: https://web.stanford.edu/ class/archive/stats/stats200/stats200.1172/Lecture24.pdf

2016

[22] [22]

Adam: A Method for Stochastic Optimization

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inInternational Conference on Learning Representations (ICLR), 2015, arXiv:1412.6980. [Online]. Available: https://arxiv.org/abs/1412.6980

work page internal anchor Pith review Pith/arXiv arXiv 2015

[23] [23]

SciPy 1.0--Fundamental Algorithms for Scientific Computing in Python

P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, İ. Polat, Y. Feng, E. W. Moore, J. VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Henriksen, ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1038/s41592-019-0686-2 2020