Pith. sign in

REVIEW 4 major objections 5 minor 96 references

Efficient Unified Caching for Accelerating Heterogeneous AI Workloads

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read IGTCache claims that one cluster-wide cache can lift cache hit ratio by 55.6% and cut average job completion time by 52.2% by detecting each stream's access pattern and granularity and switching prefetch, eviction, and allocation policies…

desk verdict A genuine, well-scoped systems contribution that adapts known caching policies per access stream using a hierarchical access tree and K-S hypothesis testing; the end-to-end wins are plausible but rest on a single testbed and an underspecified statistical assumption about dataset size c. read the letter →

arxiv 2506.12370 v1 pith:3PRLJP2S submitted 2025-06-14 cs.DC cs.LG

Tianze Wang , Yifei Liu , Chen Chen , Pengfei Zuo , Jiawei Zhang , Qizhen Weng , Yin Chen , Zhenhua Han
show 3 more authors
Jieru Zhao Quan Chen Minyi Guo
This is my paper
classification cs.DCcs.LG
keywords unifiedcacheAIclustersaccesspatternclassificationK-StestprefetchingevictionallocationStreamTree
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

IGTCache sets out to show that one cache shared by every workload in an AI cluster can beat caches tuned for a single workload type. It organizes recent remote accesses into an AccessStreamTree, a prefix tree whose nodes are access streams at block, file, or directory granularity, and runs a K-S hypothesis test on the spatial gaps between consecutive accesses to label each stream sequential, random, or skewed. Prefetching, eviction, and cache-space allocation are then chosen per stream: sequential streams prefetch ahead and evict eagerly, random streams pin data uniformly and can prefetch whole datasets, skewed streams use LRU and no prefetching. In an 18-job mixed testbed the paper measures a 55.6% higher cache hit ratio than existing caching frameworks and a 52.2% lower average job completion time, with computation overhead of 0.36% of average I/O time. The paper frames the contribution as enabling adaptivity among known workload-specific policies rather than inventing a new caching algorithm.

What carries the argument

The AccessStreamTree, a hierarchical index of recent accesses built by prefix-matching each request's path, with one AccessStream node per granularity, is what lets pattern detection occur at block, file, or directory level. The second carrier is the K-S hypothesis test, which compares the empirical spatial-gap distribution to the triangular CDF $F(k)=2k/(c-1)-k(k+1)/(c(c-1))$ to call a stream random, with other shapes treated as skewed and small gaps as sequential. The third is the marginal-benefit metric $B$ (for example, $B=1/(g\cdot n)$ for random streams), a common currency that lets cache space migrate between streams every 60 seconds.

What would settle it

Generate a stream that samples a dataset of known size $c$ uniformly with replacement and compute the empirical spatial-gap distribution over 100 consecutive accesses; if the K-S test at $\alpha=0.01$ rejects the triangular reference $F(k)=2k/(c-1)-k(k+1)/(c(c-1))$ while the stream is genuinely random, the paper's random-pattern model is falsified and the policies keyed to it would be misconfigured.

Watch

Extended reading notes

Core claim

IGTCache's central claim is that one shared cache can serve the full mixture of preprocessing, training, and inference workloads because each access stream is classified at runtime and given the policy that fits its pattern and granularity. Recent accesses are folded into an AccessStreamTree, whose nodes correspond to directory, file, or block granularity; each non-trivial stream is classified by a K-S test comparing the empirical spatial-gap distribution against the triangular reference $P(Z=k)=2(c-k)/(c(c-1))$ that uniform sampling without replacement produces. Sequential streams get sequential prefetch and eager eviction, random streams get dataset prefetching and uniform pinning, and skewed streams get LRU and no prefetching, while a marginal-benefit metric shifts cache space among streams. On an 18-job mixed workload the paper measures a 55.6% higher cache hit ratio than existing caching frameworks and a 52.2% lower average job completion time, with 0.36% average I/O-time computation overhead.

Load-bearing premise

The pattern detector assumes that a random workload touches every data item exactly once per pass, that the cache knows the dataset's total item count, and that random and non-random accesses never mix in one stream; if a data loader samples with replacement, shards data, or mixes patterns, the triangular gap distribution breaks and the K-S test can pick the wrong policy class for the stream.

Editorial extensions

If this is right

  • A cluster can run a single shared cache instead of one cache per workload, because each AccessStream keeps its policy decisions isolated even though the cache space is shared.
  • Pattern changes are handled automatically: if a dataset shifts from training (random) to testing (sequential), the same AccessStream is reclassified and its prefetching and eviction policies switch accordingly.
  • The reported gain is largest when cache is scarce, with a 55.6% hit-ratio improvement at 35% dataset-size cache, and gains persist even at 100% cache size because prefetching removes compulsory misses.
  • Random training streams receive statistical dataset prefetching, which alone reduced first-epoch completion time by 6.8% in the paper's measurement.
  • Adaptive TTL eviction releases cache space of finished jobs much sooner (86 seconds versus 600 seconds in the reported experiment), so surviving jobs see higher throughput.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper's testbed, the classifier consumes only access sequences and item counts, so the same machinery could be wrapped around page caches, object-store gateways, or data-loading pipelines to test whether pattern-switched policies reproduce the reported gains elsewhere.
  • The uniform-without-replacement reference distribution will need generalization before the method transfers to data loaders that sample with replacement, shard across workers, or repeat items within an epoch; an occupancy-based or beta-binomial gap model is a concrete candidate.
  • Aggregate JCT and hit-ratio numbers likely hide heterogeneous per-job effects, so a percentile breakdown would tell operators which workload class receives most of the 52.2% job-completion-time reduction and whether any class regresses.
  • Because every downstream policy is determined by a single K-S verdict, an online calibration of the significance level per stream, rather than a fixed 0.01, would be a natural robustness follow-up.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents IGTCache, a unified cache layer for AI clusters that serves heterogeneous workloads (data preprocessing, training, inference) in a single shared cache pool. IGTCache organizes recent accesses into an AccessStreamTree, whose nodes represent access streams at directory, file, or block granularity. For each non-trivial stream, it uses a Kolmogorov–Smirnov test against a triangular gap distribution to classify the access pattern as sequential, random, or skewed. Based on the classification, it adaptively selects prefetching, eviction, and cache-space-allocation policies per stream. The experimental section reports a 55.6% cache hit ratio improvement and a 52.2% average job completion time reduction over JuiceFS on a hand-assembled 18-job mixed workload, with computation overhead of 0.36% of average I/O time. The paper also includes micro-benchmarks for prefetching, eviction, allocation, and sensitivity analysis of the K-S test parameters.

Significance. If the claims hold, the paper offers a practical recipe for making a single cache pool adapt to the heterogeneous access patterns and storage granularities found in modern AI clusters, without requiring application code changes. The AccessStreamTree abstraction is clearly presented, the implementation on JuiceFS is concrete, and the inclusion of sensitivity analysis for the significance level and observation window is a genuine robustness check. The paper also makes falsifiable predictions, such as the adaptive TTL behavior evaluated in §5.3. However, the statistical classifier underlying the three adaptive policies rests on assumptions that are only partially validated: it requires knowledge of the total item count c, assumes uniform sampling without replacement, and applies continuous K-S critical values to a discrete variable. These issues, together with the single-run nature of the headline end-to-end results, prevent the current manuscript from fully supporting its central claims.

major comments (4)
  1. [§3.2, Eq. (1)] The reference distribution for the K-S test, F(k) = 2k/(c-1) - k(k+1)/(c(c-1)), requires the total item count c of the dataset, but the paper never states how IGTCache obtains c for an AccessStream at file or directory granularity. The implementation section (§4) describes the FUSE hook and tree updates but does not mention any metadata lookup that would provide c. If c is unknown, the test statistic cannot be computed; if it is estimated from the observation window, the estimate may be biased for small windows. Because the resulting classification drives prefetching, eviction, and allocation in §3.3, this is a load-bearing omission. Please specify the source of c and provide a sensitivity analysis of the classifier to errors in c.
  2. [§3.2, footnote 6; Table 3] The triangular gap model assumes uniform sampling without replacement from the full dataset. Two failure modes are unaddressed: (i) distributed data loaders shard the dataset, so the merged stream combines independent uniform streams over shards, yielding a different gap distribution; (ii) epoch boundaries cause repeated samples, violating the without-replacement assumption. Footnote 6 explicitly excludes mixed random/non-random streams, but the workload suite in Table 3 includes job 18, whose pattern is listed as 'Sequential and Random' (multimodal finetuning), and other workloads may share datasets. The paper should either demonstrate that misclassification is rare under these conditions (e.g., by testing on sharded loaders and multi-epoch runs) or qualify the headline claims accordingly.
  3. [§5.1] The headline numbers—55.6% CHR improvement and 52.2% JCT reduction—are point estimates from a single run of one hand-assembled 18-job mix against one baseline (JuiceFS). No confidence intervals, multiple seeds, or workload-composition sensitivity are reported. The micro-benchmarks in §5.2–§5.4 compare against stronger baselines, but the end-to-end claim is the paper's central result. Several independent runs (e.g., different job arrival intervals and dataset orderings) are needed to establish that the gains are not an artifact of a particular mix.
  4. [§3.2, K-S test application] The K-S test is applied to a discrete variable (spatial gaps Z with integer values), yet the critical values D_α are those of the continuous Kolmogorov distribution. With ties, the standard test is conservative, which biases towards rejecting the random-pattern null hypothesis. The practical consequence is that truly random streams may be classified as skewed, switching eviction from uniform caching to LRU and disabling statistical prefetching—exactly the actions that drive the §5.1 gains. The authors should use a discrete K-S test or provide a correction and quantify the impact.
minor comments (5)
  1. [Abstract and §5.1] The abstract and §5.1 describe JuiceFS as a 'state-of-the-art' caching framework, but Alluxio is only mentioned, not evaluated; the claim of improvement 'over state-of-the-art caching frameworks' is broader than the experimental evidence.
  2. [Table 3] For job 18 (LLaVa multimodal finetuning), the pattern is listed as 'Sequential and Random'; given footnote 6's exclusion of mixed random and non-random streams, it is unclear whether the AccessStreamTree separates the text and image accesses into different streams by path. Please clarify how this workload is handled.
  3. [§3.3] For random patterns, the marginal benefit is defined as B = 1/(g·n), where g is the 'inter-access temporal gap'; the text does not define how g is measured for a stream with multiple concurrent workers, and the assumption of a single constant g is not justified.
  4. [§4] The hard limit of 10,000 nodes in the AccessStreamTree, with excess nodes removed by LRU, can evict the state of a non-trivial AccessStream and reset its classification; the effect on ongoing policy adaptation is not discussed.
  5. [§5.5] Figures 14 and 15 report classification accuracy but do not state the numerical values in the text; the claim that α=0.01 is 'sufficiently good' would be easier to evaluate with stated numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: pattern classification is an operational hypothesis test, policy gains are measured end-to-end, and the sole overlapping-author citation is independent support.

full rationale

The claimed derivation is not circular. IGTCache's pattern recognition (Section 3.2) tests observed spatial-gap samples against the triangular reference distribution P(Z=k)=2(c-k)/(c(c-1)) derived from the uniform-without-replacement model; the reference distribution is a mathematical consequence of the assumed random-access model, not a fit to the workloads, and Section 5.5 checks the classifier against independently labeled workload types over 100 trials per pattern. The adaptive policies (Section 3.3) are selected from the recognized pattern by construction, but the headline JCT/CHR results (Section 5.1) are measured end-to-end on the 18-job testbed and each component is ablated in Sections 5.2-5.4 against external baselines (stride, SFP, LRU, FIFO, ARC, uniform caching, Quiver, Fluid). No parameter is fitted to the headline metrics: alpha=0.01, window=100, prefetch depth=4, and BufferWindow=100 are stated defaults, and the marginal benefit B is measured via a ghost cache or computed from dataset size n and access gap g rather than tuned to JCT. The one overlapping-author citation, Silod [87] for the uniform-caching eviction policy, is independent published work and is also supported by the non-overlapping citation [58], so it is not load-bearing self-citation. Footnote 6's exclusion of mixed random/non-random streams and the paper's silence on how c in Eq. (1) is obtained are correctness/generality risks, not circular reductions: nothing in the paper defines c in terms of the classification outcome or equates a prediction with a fitted value.

Assumptions & free parameters 8 free parameters · 6 assumptions · 1 invented entities

The central claim rests on the pattern-classification model of Section 3.2: Eq. (1) plus known c, uniform-without-replacement sampling, and the no-mixed-pattern exclusion (footnote 6). The per-pattern policy mappings (eager eviction, uniform pinning, LRU, no skewed prefetching) are carried over from prior work and two motivation experiments rather than derived. Eight hand-set hyperparameters (alpha=0.01, window=100, prefetch depth 4, f_p=0.8, BufferWindow w=100, TTL base 60 s, shift round 60 s / 640 MB, minimum share 640 MB) are not fitted to the headline result in a formal sense, but the TTL base time and shift round coincide with the testbed's job inter-arrival expectation beta=60 s, so the end-to-end numbers are sensitive to these choices. No new physical entities are postulated; the AccessStreamTree and CacheManageUnit are implemented software abstractions with measured overhead.

free parameters (8)
  • K-S significance level alpha = 0.01
    Hand-chosen significance level for pattern classification; Section 5.5 reports sensitivity (Figure 14) showing 0.01 is in general better than 0.05, so it is tuned but its robustness is checked.
  • Observation window size = 100
    Trigger for classifying an AccessStream as non-trivial; Section 3.2 sets it to 100, and Section 5.5 shows accuracy improves up to 100 with little gain at 1000.
  • Sequential prefetch depth N = 4
    Fixed prefetch depth for sequential streams (Section 3.3); no sensitivity analysis is reported for this value.
  • Hot-unit prefetch threshold f_p = 0.8
    Threshold on access probability for selective hierarchical prefetching (Section 3.3); no sensitivity analysis is reported.
  • BufferWindow size w = 100
    Ghost-cache buffer size for marginal benefit estimation in skewed streams (Section 3.3); the value is stated without justification.
  • TTL base time = 60 s
    Added to the statistically estimated TTL to avoid accidental eviction (Section 3.3); its value equals the job inter-arrival expectation beta=60 s in the testbed (Section 5.1), a suspicious alignment for the Figure 11 result.
  • Cache shifting round and amount = 640 MB per 60 s round
    Configuration of the allocation migration loop (Section 4); 640 MB also serves as the minimum cache share per dataset.
  • Dataset-prefetch hit-ratio threshold = not specified in text
    Statistical prefetching for random patterns triggers when the expected cache hit ratio exceeds a threshold (Section 3.3), but neither the estimator nor the threshold value is given.
assumptions (6)
  • domain assumption Random-access streams sample item indices uniformly without replacement within an epoch, so consecutive-access gaps follow P(Z=k)=2(c-k)/(c(c-1)).
    This is the null distribution for the K-S test in Section 3.2 (Eq. 1). It breaks under resampling, sharded or multi-worker samplers, and epoch-boundary repeats, none of which are discussed.
  • domain assumption IGTCache knows the dataset item count c required by the reference CDF of Eq. (1).
    F(k)=2k/(c-1)-k(k+1)/(c(c-1)) is parameterized by c; the paper never states how c is discovered or refreshed at runtime, though some datasets in Table 1 have well-known sizes.
  • ad hoc to paper No AccessStream mixes random and non-random access patterns.
    Admitted in footnote 6 (Section 3.2): mixed patterns threaten data consistency and are rare in practice. This excludes the shared-dataset scenario cited as motivation in Section 2.1.
  • domain assumption Continuous-distribution K-S critical values remain valid for the discrete triangular CDF.
    Section 3.2 applies the standard K-S test to a lattice distribution without the usual discreteness correction or Monte Carlo critical values; the approximation is not mentioned.
  • domain assumption The per-pattern policy mappings are the right ones: eager eviction for sequential, uniform pinning for random, LRU and no prefetching for skewed.
    Adopted in Section 3.3 from the two motivation experiments in Section 2.2 and prior work [58,87]; there is no ablation showing these mappings beat alternatives within each pattern class.
  • domain assumption Temporal gaps between consecutive accesses of a random-pattern stream are normally distributed.
    Adaptive TTL (Section 3.3) fits a normal distribution to measured temporal gaps and sets TTL as a significance-thresholded maximum; the distributional claim is asserted without supporting evidence.
invented entities (1)
  • AccessStreamTree (with per-node AccessStream and per-stream CacheManageUnit) independent evidence
    purpose: Hierarchical history of recent accesses organized by path prefix at block, file, and directory granularity; each node is an independent unit for pattern recognition, policy customization, and cache-space isolation (Section 3.1, Section 4).
    These are software abstractions contributed by the paper, not postulated physical entities. The falsifiable handles are internal but concrete: pattern-recognition accuracy versus alpha and window size (Section 5.5), O(log N) update with measured 47.6 microseconds per-I/O overhead and 73.2 MB memory at 10,000 nodes (Section 5.5), and the end-to-end CHR/JCT measurements (Section 5.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Unified Caching for Accelerating Heterogeneous AI Workloads." pith.science (2026). https://pith.science/paper/3PRLJP2S

@misc{pith2026250612370,
  author       = {Pith},
  title        = {Pith review of: Efficient Unified Caching for Accelerating Heterogeneous AI Workloads},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3PRLJP2S}},
  note         = {Machine review of arXiv:2506.12370}
}
read the original abstract

Modern AI clusters, which host diverse workloads like data pre-processing, training and inference, often store the large-volume data in cloud storage and employ caching frameworks to facilitate remote data access. To avoid code-intrusion complexity and minimize cache space wastage, it is desirable to maintain a unified cache shared by all the workloads. However, existing cache management strategies, designed for specific workloads, struggle to handle the heterogeneous AI workloads in a cluster -- which usually exhibit heterogeneous access patterns and item storage granularities. In this paper, we propose IGTCache, a unified, high-efficacy cache for modern AI clusters. IGTCache leverages a hierarchical access abstraction, AccessStreamTree, to organize the recent data accesses in a tree structure, facilitating access pattern detection at various granularities. Using this abstraction, IGTCache applies hypothesis testing to categorize data access patterns as sequential, random, or skewed. Based on these detected access patterns and granularities, IGTCache tailors optimal cache management strategies including prefetching, eviction, and space allocation accordingly. Experimental results show that IGTCache increases the cache hit ratio by 55.6% over state-of-the-art caching frameworks, reducing the overall job completion time by 52.2%.

Figures

Figures reproduced from arXiv: 2506.12370 by the authors.

Figure 1
Figure 1. Workloads, cache and storage in AI clusters. Eviction chooses the data items to kick out when there is no idle space [58]. Space allocation means to set the amount of cache space allowed to be used by each workload [39, 40, 49]. Those policies are crucial to the overall caching performance of the AI clusters. The need for a unified, high-performance cache for AI clusters. For high resource utilization, AI clusters a… view at source ↗
Figure 2
Figure 2. (a) Block-level prefetching (B-P) speeds up pre￾processing on Bookcorpus but fails to benefit ImageNet test￾ing, whereas file-level prefetching (F-P) has the opposite ef￾fect. The times are normalized relative to the completion time without prefetching. (b) LRU eviction outperforms uniform￾caching policy for the RAG workload but performs worse for ImageNet training. Heterogeneity of data access patterns (sequential,… view at source ↗
Figure 3
Figure 3. Architecture of IGTCache. Challenges. While promising, realizing automatic cache management in practice is nonetheless a non-trivial task, and the challenges are as follows: Challenge 1: A production cache often concurrently serves requests from multiple AI workloads. While the data access pattern of an AI workload is relatively clear, multiple access sequences from different workloads may mix up with each other, ma… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: The structure of an AccessStreamTree. and accurate pattern recognition, IGTCache introduces a hy￾pothesis testing method to identify the access pattern for each workload at runtime, addressing Challenge 2 (§3.2). Third, for high efficacy, it customizes the prefetching,…
Figure 6
Figure 6. Figure 6: Distributions of spatial gaps between any two consecutive accesses for workloads with different patterns. (a) Triangular shape for a training workload with random pattern. (b) Non-triangular shape for a query-based data augmentation workload [28] with skewed pattern. G…
Figure 7
Figure 7. Figure 7: Example of hierarchical prefetching over an ocean atmosphere dataset (ICOADS [7]), with semantic information in path names. The directory 202209/ contains all data col￾lected in September 2022 (a total of 10 csv files), and the file 10 100 00 110.csv stores the data fo…
Figure 8
Figure 8. Figure 8: JCT and CHR in end-to-end evaluations. All Job Job Job Job Job Job Job 0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 Normalized JCT IGTCache No-Prefetch Stride JuiceFS SFP All Job Job Job Job Job Job Job 0 20 40 60 80 100 CHR (%) [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Performance of different prefetching schemes. 5.2 Prefetching Performance Setup. To evaluate IGTCache performance in prefetching, we select a subset of jobs that are sensitive to the prefetch￾ing effect (job IDs marked in [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 11
Figure 11. Figure 11: After job-○9 finishes at 60s, the eviction of its ImageNet dataset starts at 146s if using adaptive TTL setup in IGTCache, and at 660s if using default TTL setup. (with ImageNet dataset), and measure the throughput of job- ○13 respectively with the default10 TTL setup…
Figure 14
Figure 14. Figure 14: Accuracy with dif￾ferent confidence 𝛼. Mean Rand. Skewed 60 65 70 75 80 85 90 95 100 Accuracy (%) size=10 size=100 size=1000 [PITH_FULL_IMAGE:figures/full_fig_p012_14.png]
Figure 13
Figure 13. Figure 13: Instantaneous marginal cache benefit and cache allocation amount of each job under IGTCache [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]
Figure 16
Figure 16. Figure 16: CHR under differ￾ent cache sizes. 10 0 10 1 10 2 10 3 10 4 10 5 10 0.0 6 0.5 1.0 Overall JCT Normalized JCT 10 0 10 1 10 2 10 3 10 4 10 5 10 0 6 50 Time ( s) Comp. per I/O 10 0 10 1 10 2 10 3 10 4 10 5 10 6 Number of Nodes 0 3 6 Memory (GB) Memory Usage [PITH_FULL_IM…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

96 extracted references · 64 canonical work pages

  1. [1]

    2024. Alluxio. https://www.alluxio.io/. Retrieved on September 12, 2024

    2024

  2. [2]

    ALLUXIO FUSE

    2024. ALLUXIO FUSE. https://docs.alluxio.io/os/user/stable/en/api/ POSIX-API.html. Retrieved on September 12, 2024

    2024

  3. [3]

    Amazon S3

    2024. Amazon S3. https://aws.amazon.com/s3/ . Retrieved on September 12, 2024

    2024

  4. [4]

    Azure Blob

    2024. Azure Blob. https://azure.microsoft.com/en-us/products/storag e/blobs. Retrieved on September 12, 2024

    2024

  5. [5]

    Data preloading in Alluxio

    2024. Data preloading in Alluxio. https://docs.alluxio.io/ee-ai/user/st able/en/feature/Cache-Preloading.html. Retrieved on September 12, 2024

    2024

  6. [6]

    Free Speech… Recognition (Linux, Windows and Mac) - vox- forge.org

    2024. Free Speech… Recognition (Linux, Windows and Mac) - vox- forge.org. https://www.voxforge.org/. Retrieved on July 31, 2024

    2024

  7. [7]

    International Comprehensive Ocean-Atmosphere Data Set (ICOADS) Release 3, Individual Observations

    2024. International Comprehensive Ocean-Atmosphere Data Set (ICOADS) Release 3, Individual Observations. https://rda.ucar.e du/datasets/dsd548000/. Retrieved on September 12, 2024

    2024

  8. [8]

    2024. JuiceFS. https://juicefs.com/en/ . Retrieved on September 12, 2024

    2024

Show all 96 references
  1. [9]

    JuiceFS cache strategies

    2024. JuiceFS cache strategies. https://juicefs.com/docs/cloud/guide /cache/#consistency-exceptions. Retrieved on September 12, 2024

    2024

  2. [10]

    JuiceFS FUSE

    2024. JuiceFS FUSE. https://juicefs.com/docs/community/fuse mou nt options/. Retrieved on September 12, 2024

    2024

  3. [11]

    Time Series Air Quality Data of India (2010-2023)

    2024. Time Series Air Quality Data of India (2010-2023). https: //www.kaggle.com/datasets/abhisheksjha/time-series-air-quality- data-of-india-2010-2023 . Retrieved on September 11, 2024

    2024

  4. [12]

    TTL (Time-to-live) setup in Alluxio

    2024. TTL (Time-to-live) setup in Alluxio. https://docs.alluxio.io/os/u ser/stable/en/core-services/Caching.html. Retrieved on September 12, 2024

    2024

  5. [13]

    TTL (Time-to-live) setup in JuiceFS

    2024. TTL (Time-to-live) setup in JuiceFS. https://juicefs.com/docs/c ommunity/guide/cache/. Retrieved on September 12, 2024

    2024

  6. [14]

    What is Object storage? https://cloud.google.com/learn/what- is-object-storage?hl=en

    2024. What is Object storage? https://cloud.google.com/learn/what- is-object-storage?hl=en. Retrieved on September 12, 2024

    2024

  7. [15]

    Configure Object Storage with Milvus Operator

    2025. Configure Object Storage with Milvus Operator. https://milvus .io/docs/object storage operator.md. Retrieved on May 16, 2025

    2025

  8. [16]

    deepseek-v3

    2025. deepseek-v3. https://ollama.com/library/deepseek-v3. Retrieved on May 16, 2025

    2025

  9. [17]

    GPT-4: Details Leaked

    2025. GPT-4: Details Leaked. https://plainswipe.com/gpt-4-details- leaked/index.html. Retrieved on January 10, 2025

    2025

  10. [18]

    GPT-4o System Card

    2025. GPT-4o System Card. https://openai.com/index/gpt-4o-system- card/. Retrieved on January 10, 2025

    2025

  11. [19]

    2025. Milvus. https://milvus.io/. Retrieved on January 10, 2025

    2025

  12. [20]

    Wikipedia (en) embedded with cohere.ai multilingual22-12 encoder

    2025. Wikipedia (en) embedded with cohere.ai multilingual22-12 encoder. https://huggingface.co/datasets/Cohere/wikipedia-22-12- en-embeddings/. Retrieved on January 10, 2025

    2025

  13. [21]

    Param Aggarwal. 2019. Fashion Product Images Dataset. https: //doi.org/10.34740/KAGGLE/DS/139630

    2019 doi

  14. [22]

    Maen M Al Assaf. 2015. Predictive Prefetching for Parallel Hybrid Storage Systems. International Journal of Communications, Network and System Sciences 8, 5 (2015), 161–180

    2015

  15. [23]

    Hasan Al Maruf and Mosharaf Chowdhury. 2020. Effectively prefetch- ing remote memory with leap. In USENIX ATC

    2020

  16. [24]

    Kolmogorov An. 1933. Sulla determinazione empirica di una legge didistribuzione. Giorn Dell’inst Ital Degli Att 4 (1933), 89–91

    1933

  17. [25]

    Sorav Bansal, Dharmendra S Modha, et al . 2004. Car: Clock with adaptive replacement.. In FAST, Vol. 4. 187–200

    2004

  18. [26]

    S¨oren Becker, Johanna Vielhaben, Marcel Ackermann, Klaus-Robert M¨uller, Sebastian Lapuschkin, and Wojciech Samek. 2024. AudioM- NIST: Exploring Explainable Artificial Intelligence for audio analysis on a simple benchmark. Journal of the Franklin Institute 361, 1 (2024), 418–428

    2024

  19. [27]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei

  20. [28]

    Yuhao Deng, Chengliang Chai, Lei Cao, Qin Yuan, Siyuan Chen, Yanrui Yu, Zhaoze Sun, Junyi Wang, Jiajun Li, Ziqi Cao, et al. 2024. LakeBench: A Benchmark for Discovering Joinable and Unionable Tables in Data Lakes. Proceedings of the VLDB Endowment 17, 8 (2024), 1925–1938

    2024

  21. [29]

    Shi Dong, Ping Wang, and Khushnood Abbas. 2021. A survey on deep learning and its applications. Computer Science Review 40 (2021), 100379

    2021

  22. [30]

    Nikoli Dryden, Roman B ¨ohringer, Tal Ben-Nun, and Torsten Hoefler

  23. [31]

    Gil Einziger, Ohad Eytan, Roy Friedman, and Benjamin Manes. 2022. Lightweight robust size aware cache management. ACM Transactions on Storage (TOS) 18, 3 (2022), 1–23

    2022

  24. [32]

    Gil Einziger, Roy Friedman, and Ben Manes. 2017. Tinylfu: A highly efficient cache admission policy. ACM Transactions on Storage (ToS) 13, 4 (2017), 1–31

    2017

  25. [33]

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. 2010. The pascal visual object classes (voc) challenge. International journal of computer vision 88 (2010), 303–338

    2010

  26. [34]

    Peter X Gao, Akshay Narayan, Sagar Karandikar, Joao Carreira, Sangjin Han, Rachit Agarwal, Sylvia Ratnasamy, and Scott Shenker

  27. [35]

    Shaleen Garg, Jian Zhang, Rekha Pitchumani, Manish Parashar, Bing Xie, and Sudarsun Kannan. 2024. CrossPrefetch: Accelerating I/O Prefetching for Modern Storage. In ACM ASPLOS

    2024

  28. [36]

    Binny S Gill and Luis Angel D Bathen. 2007. AMP: Adaptive Multi- stream Prefetching in a Shared Cache.. In USENIX FAST

    2007

  29. [37]

    Xiangyang Gou, Long He, Yinda Zhang, Ke Wang, Xilai Liu, Tong Yang, Yi Wang, and Bin Cui. 2020. Sliding sketches: A framework using time zones for data stream processing in sliding windows. In PAKDD

    2020

  30. [38]

    Jim Griffioen and Randy Appleton. 1994. Reducing File System Latency using a Predictive Approach.. In USENIX summer. 197–207

    1994

  31. [39]

    Rong Gu, Simian Li, Haipeng Dai, Hancheng Wang, Yili Luo, Bin Fan, Ran Ben Basat, Ke Wang, Zhenyu Song, Shouwei Chen, et al . 2023. Adaptive online cache capacity optimization via lightweight working set size estimation at scale. In USENIX ATC

    2023

  32. [40]

    Rong Gu, Kai Zhang, Zhihao Xu, Yang Che, Bin Fan, Haojun Hou, Haipeng Dai, Li Yi, Yu Ding, Guihai Chen, et al. 2022. Fluid: Dataset abstraction and elastic acceleration for cloud-native deep learning training jobs. In IEEE ICDE. 13 Wang et al

    2022

  33. [41]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In IEEE CVPR

    2016

  34. [42]

    Virajith Jalaparti, Chris Douglas, Mainak Ghosh, Ashvin Agrawal, Avrilia Floratou, Srikanth Kandula, Ishai Menache, Joseph Seffi Naor, and Sriram Rao. 2018. Netco: Cache and i/o management for analytics over disaggregated stores. In ACM SoCC

    2018

  35. [43]

    Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer

  36. [44]

    Aarati Kakaraparthy, Abhay Venkatesh, Amar Phanishayee, and Shiv- aram Venkataraman. 2019. The case for unifying data loading in machine learning clusters. In 11th USENIX Workshop on Hot Topics in Cloud Computing (HotCloud 19)

    2019

  37. [45]

    Redwan Ibne Seraj Khan, Ahmad Hossein Yazdani, Yuqi Fu, Arnab K Paul, Bo Ji, Xun Jian, Yue Cheng, and Ali R Butt. 2023. SHADE: Enable Fundamental Cacheability for Distributed Deep Learning Training. In USENIX FAST

    2023

  38. [46]

    Nicholas Krichevsky, Renee St Louis, and Tian Guo. 2021. Quantifying and improving performance of distributed deep learning with cloud storage. In IEEE IC2E

    2021

  39. [47]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Im- agenet classification with deep convolutional neural networks. Ad- vances in neural information processing systems 25 (2012)

    2012

  40. [48]

    Thomas M Kroeger and Darrell Long. 2001. Design and implementa- tion of a predictive file prefetching algorithm. (2001)

    2001

  41. [49]

    Abhishek Vijaya Kumar and Muthian Sivathanu. 2020. Quiver: An informed storage cache for deep learning. In USENIX FAST

    2020

  42. [50]

    Hui Lei and Dan Duchamp. 1997. An analytical approach to file prefetching

    1997

  43. [51]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K ¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt ¨aschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processi...

    2020

  44. [52]

    Zhenmin Li, Zhifeng Chen, Sudarshan M Srinivasan, Yuanyuan Zhou, et al. 2004. C-miner: Mining block correlations in storage systems.. In USENIX FAST

    2004

  45. [53]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Per- ona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. 2014. Mi- crosoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, ...

    2014

  46. [54]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al . 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  47. [55]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual Instruction Tuning

    2023

  48. [56]

    Nimrod Megiddo and Dharmendra S Modha. 2003. ARC: A Self-Tuning, low overhead replacement cache. In USENIX FAST

    2003

  49. [57]

    Madalin Mihailescu, Gokul Soundararajan, and Cristiana Amza. 2013. {MixApart}: Decoupled Analytics for Shared Storage Systems. In 11th USENIX Conference on File and Storage Technologies (FAST 13) . 133–146

    2013

  50. [58]

    Jayashree Mohan, Amar Phanishayee, Ashish Raniwala, and Vijay Chidambaram. 2020. Analyzing and mitigating data stalls in DNN training. arXiv preprint arXiv:2007.06775 (2020)

    2020 arXiv

  51. [59]

    Djob Mvondo, Mathieu Bacou, Kevin Nguetchouang, Lucien Ngale, St´ephane Pouget, Josiane Kouam, Renaud Lachaize, Jinho Hwang, Tim Wood, Daniel Hagimont, et al. 2021. OFC: an opportunistic caching system for FaaS platforms. In ACM EuroSys

    2021

  52. [60]

    Elizabeth J O’neil, Patrick E O’neil, and Gerhard Weikum. 1993. The LRU-K page replacement algorithm for database disk buffering. Acm Sigmod Record 22, 2 (1993), 297–306

    1993

  53. [61]

    Ray Perrault and Jack Clark. 2024. Artificial intelligence index report

    2024

  54. [62]

    Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. 2015. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to- sentence models. In Proceedings of the IEEE international conference on computer...

    2015

  55. [63]

    Moinuddin K Qureshi, Aamer Jaleel, Yale N Patt, Simon C Steely, and Joel Emer. 2007. Adaptive insertion policies for high performance caching. ACM SIGARCH Computer Architecture News 35, 2 (2007), 381–391

    2007

  56. [64]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang

  57. [65]

    Adib Bin Rashid and Ashfakul Karim Kausik. 2024. AI revolutioniz- ing industries worldwide: A comprehensive overview of its diverse applications. Hybrid Advances (2024), 100277

    2024

  58. [66]

    Susmita Ray. 2019. A quick review of machine learning algorithms. In 2019 International conference on machine learning, big data, cloud and parallel computing (COMITCon). IEEE, 35–39

    2019

  59. [67]

    Zujie Ren, Jian Wan, Weisong Shi, Xianghua Xu, and Min Zhou. 2013. Workload analysis, implications, and optimization on a production hadoop cluster: A case study on taobao. IEEE Transactions on Services Computing 7, 2 (2013), 307–321

    2013

  60. [68]

    Francisco Romero, Gohar Irfan Chaudhry, ´I˜nigo Goiri, Pragna Gopa, Paul Batum, Neeraja J Yadwadkar, Rodrigo Fonseca, Christos Kozyrakis, and Ricardo Bianchini. 2021. FaaT: A transparent auto- scaling cache for serverless applications. In ACM SoCC

    2021

  61. [69]

    arXiv preprint arXiv:1606.05250 (2016)

    Squad: 100,000+ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250 (2016)

    2016 arXiv

  62. [70]

    Sambhav Satija, Chenhao Ye, Ranjitha Kosgi, Aditya Jain, Romit Kankaria, Yiwei Chen, Andrea C Arpaci-Dusseau, Remzi H Arpaci- Dusseau, and Kiran Srinivasan. 2025. Cloudscape: A Study of Storage Services in Modern Cloud Architectures. In 23rd USENIX Conference on File and Stora...

    2025

  63. [71]

    Karen Simonyan and Andrew Zisserman. 2014. Very deep convo- lutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  64. [72]

    Yannis Smaragdakis, Scott Kaplan, and Paul Wilson. 1999. EELRU: simple and effective adaptive page replacement. ACM SIGMETRICS Performance Evaluation Review 27, 1 (1999), 122–133

    1999

  65. [73]

    Nickolay Smirnov. 1948. Table for estimating the goodness of fit of empirical distributions. The annals of mathematical statistics 19, 2 (1948), 279–281

    1948

  66. [74]

    Ayodeji Olalekan Salau and Shruti Jain. 2019. Feature extraction: a survey of the types, techniques, applications. In IEEE ICSC

    2019

  67. [75]

    Aidan Toner-Rodgers. 2024. Artificial intelligence, scientific discovery, and product innovation. arXiv preprint arXiv:2412.17866 (2024)

    2024 arXiv

  68. [76]

    Han Wang, Longfei Luo, Liang Shi, Changlong Li, Chun Jason Xue, Qingfeng Zhuge, and Edwin H-M Sha. 2021. SFP: Smart file-aware prefetching for flash based storage systems. In ACM GLSVLSI

    2021

  69. [77]

    Fengguang Wu, Hongsheng Xi, Jun Li, and Nanhai Zou. 2007. Linux readahead: less tricks for more. InProceedings of the Linux Symposium, Vol. 2. Citeseer, 273–284

    2007

  70. [78]

    Yi Xu, Jiandong Ding, Lu Zhang, and Shuigeng Zhou. 2021. Dp-ssl: Towards robust semi-supervised learning with a few labeled samples. Advances in Neural Information Processing Systems 34 (2021), 15895– 15907

    2021

  71. [79]

    Gokul Soundararajan, Madalin Mihailescu, and Cristiana Amza. 2008. Context-Aware Prefetching at the Storage Server. InUSENIX ATC

    2008

  72. [80]

    Juncheng Yang, Yao Yue, and KV Rashmi. 2021. A large-scale analysis of hundreds of in-memory key-value cache clusters at twitter. ACM Transactions on Storage (TOS) 17, 3 (2021), 1–35. 14 Efficient Unified Caching for Accelerating Heterogeneous AI Workloads

    2021

  73. [81]

    Juncheng Yang, Yazhuo Zhang, Ziyue Qiu, Yao Yue, and Rashmi Vinayak. 2023. FIFO queues are all you need for cache eviction. In ACM SOSP

    2023

  74. [82]

    Anil Yelam. 2022. Systems for memory disaggregation: challenges & opportunities. arXiv preprint arXiv:2202.02223 (2022)

    2022 arXiv

  75. [83]

    Fuxun Yu, Di Wang, Longfei Shangguan, Minjia Zhang, Chenchen Liu, and Xiang Chen. 2022. A survey of multi-tenant deep learning inference on gpu. arXiv preprint arXiv:2203.09040 (2022)

    2022 arXiv

  76. [84]

    Juncheng Yang, Reza Karimi, Trausti Sæmundsson, Avani Wildani, and Ymir Vigfusson. 2017. Mithril: mining sporadic associations for cache prefetching. In ACM SoCC

    2017

  77. [85]

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Vic- toria Lin, et al . 2022. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)

    2022 arXiv

  78. [86]

    Yazhuo Zhang, Juncheng Yang, Yao Yue, Ymir Vigfusson, and KV Rashmi. 2024. SIEVE is Simpler than LRU: an Efficient Turn-Key Eviction Algorithm for Web Caches. In USENIX NSDI

    2024

  79. [87]

    Hanyu Zhao, Zhenhua Han, Zhi Yang, Quanlu Zhang, Mingxia Li, Fan Yang, Qianxi Zhang, Binyang Li, Yuqing Yang, Lili Qiu, et al . 2023. Silod: A co-design of caching and scheduling for deep learning clusters. In ACM EuroSys

    2023

  80. [88]

    Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio Torralba, and Aude Oliva. 2014. Learning deep features for scene recognition using places database. NeurIPS (2014)

    2014

  81. [89]

    Di Zhang, Monish Soundar Raj, Bing Xie, Sheng Di, and Dong Dai

  82. [90]

    Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning Books and Movies: Towards Story-Like Visual Explanations by Watching Movies and Reading Books. In IEEE ICCV. 15

    2015

  83. [95]

    Zixuan Zhou, Xuefei Ning, Ke Hong, Tianyu Fu, Jiaming Xu, Shiyao Li, Yuming Lou, Luning Wang, Zhihang Yuan, Xiuhong Li, et al. 2024. A survey on efficient inference for large language models. arXiv preprint arXiv:2404.14294 (2024)

    2024 arXiv

  84. [2009]

    In IEEE CVPR

    Imagenet: A large-scale hierarchical image database. In IEEE CVPR

  85. [2016]

    In USENIX OSDI

    Network requirements for resource disaggregation. In USENIX OSDI

  86. [2017]

    arXiv preprint arXiv:1705.03551 (2017)

    Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551 (2017)

    2017 arXiv

  87. [2021]

    In IEEE SC

    Clairvoyant prefetching for distributed machine learning I/O. In IEEE SC

  88. [2024]

    In 2024 IEEE International Parallel and Distributed Processing Symposium (IPDPS)

    Cross-system analysis of job characterization and scheduling in large-scale computing clusters. In 2024 IEEE International Parallel and Distributed Processing Symposium (IPDPS) . IEEE, 716–727

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.