A Diffusion-Based Framework for Configurable and Realistic Multi-Storage Trace Generation

Jinhyung Koo; Jongho Park; Junyoung Lee; Seohyun Kim; Sungjin Lee; Yeseong Kim

REVIEW 4 major objections 4 minor 5 references

A Diffusion-Based Framework for Configurable and Realistic Multi-Storage Trace Generation

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A diffusion-based framework, DiTTO, generates realistic multi-device storage traces, matching user-specified read/write ratios within 8 percent average error and device utilization patterns within 2 percent.

desk verdict Plausible new diffusion-based approach to multi-device storage trace generation, but the missing image-to-event decoding means the headline 8% config error supports conditioning, not realism. read the letter →

arxiv 2509.01919 v1 pith:P53QMHGA submitted 2025-09-02 cs.CV cs.PF

Seohyun Kim , Junyoung Lee , Jongho Park , Jinhyung Koo , Sungjin Lee , Yeseong Kim This is my paper · ORCID

classification cs.CVcs.PF

keywords diffusionmodelstoragetracegenerationmulti-deviceworkloadconditioningcontrastivelearningoutpaintingsyntheticsparsity-awaretraining

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

DiTTO is a diffusion-based generative framework for producing synthetic storage workloads — the streams of read and write operations that distributed storage systems such as RAID arrays and CephFS must serve. Real workload traces are costly to collect and raise privacy concerns, while template-based synthetic generators miss the complex, evolving behavior of real systems; DiTTO argues that a conditional diffusion model can fill that gap with traces that are both realistic and precisely controllable. The framework rasterizes access logs into multi-channel images (time on one axis, storage devices on the other, separate channels for reads and writes), conditions generation on quantitative workload parameters through a contrastively trained embedding called CHIP, and uses outpainting to extend traces to arbitrary lengths. The reported result: generated traces follow user-specified read/write ratios within 8% average error and device utilization within 2%, while remaining diverse and structurally similar to real traces.

What carries the argument

The load-bearing mechanism is a conditional diffusion pipeline over an image-like encoding of the trace. Raw logs (timestamps, operation type, device ID) are rasterized into a grid: the x-axis is normalized time binned into fixed intervals, the y-axis indexes storage devices, and separate channels mark read versus write presence. Because storage workloads are extremely sparse, local Gaussian smoothing around each event creates intensity gradients that make the structure learnable; this sparsity-aware treatment is essential to preserving bursty access patterns. CHIP — Contrastive Hyperconfiguration-Image Pretraining — is the conditioning bridge: it contrastively aligns embeddings of user-supp

What would settle it

Take a real trace and a DiTTO-generated trace matched to the same configuration, replay both through the same storage emulator or stack, and compare inter-arrival-time distributions, queue depths, cache hit ratios, and tail latencies. If generated traces meet their read/write and utilization targets but their inter-arrival distributions lack the heavy-tailed burstiness of the originals, or replay outcomes diverge, the fidelity claim collapses. A direct check on the pipeline alone: extract discrete events from a generated image and compare burst-length and inter-arrival statistics against the t

Watch

Extended reading notes

Core claim

The paper's central claim: image-diffusion techniques can generate high-fidelity, precisely controllable multi-device storage traces, and DiTTO is the first diffusion-based trace generator. Three design choices carry the argument: rasterizing access logs into image-like grids (time by device, with separate read/write channels) so a U-Net captures temporal dynamics and cross-device correlations; CHIP, a contrastive-learning mechanism that aligns embeddings of numeric workload configurations with trace-image embeddings so generation obeys explicit quantitative constraints; and outpainting, which extends traces arbitrarily in time while keeping bursts and coordinated device behavior consistent.

Load-bearing premise

The framework depends on the assumption that converting discrete timestamped events to a fixed-resolution image grid, with Gaussian blur around each event, preserves the fine-grained timing and burst structure that storage workloads need — and that the smoothed image can then be converted back into discrete read/write events; the paper validates neither direction of that conversion.

Editorial extensions

If this is right

Storage-system researchers can synthesize large, privacy-free test workloads whose read/write ratios and device utilization are set by the user, substituting for traces that are costly or impossible to collect.
Trace length no longer depends on collected data: outpainting extends generation arbitrarily, keeping periodic bursts and coordinated device accesses consistent across segment boundaries.
Un-conditioned properties carry through: generated traces separate by spatial-locality cluster even though locality is never part of the conditioning, indicating the representation retains real workload structure.
The reported accuracy (under 8% read/write-ratio error and under 2% utilization error across 50 generated traces) means configuration targets can be treated as reliable experimental controls in workload studies.

Reading between the lines

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

If the claims hold, storage benchmarking can treat configuration-matching traces as experimental controls — but the paper does not yet show that generated traces reproduce system-level outcomes (cache hit rates, tail latencies) of real traces matched on the same configuration; replay-based validation is the natural next test.
The paper never specifies how continuous image intensities are binarized back into discrete read/write events; until that inverse mapping is described and checked against inter-arrival-time distributions, the discrete trace artifact has an unproven step.
The CHIP conditioning idea is workload-agnostic: the same contrastive alignment of numeric specifications with an image-like rasterization could generate other structured event streams, such as network traffic or user session logs, with quantitative targets.
Spatial locality emerging without explicit conditioning hints that the latent space organizes real workload archetypes; inverting the embedding could classify unknown workloads or flag anomalous access patterns, applications the paper does not pursue.

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.

Desk Editor's Note

read the letter

The short version: this is a sensible first pass at using diffusion models for storage trace generation, and the CHIP conditioning trick plus outpainting makes it more interesting than a toy. But the paper as written validates an image generator, not a storage trace generator. There is no description of how the smoothed, time-binned image gets converted back into timestamped read/write events, so the 8% and 2% error numbers measure how well the generated image matches the conditioning vector, not whether the result is usable by a storage simulator.

The application is new relative to what I know of the storage trace synthesis literature. The sparsity-aware Gaussian augmentation is a practical answer to a real problem, and the outpainting mechanism is a reasonable way to get long traces without retraining. The three-stage pipeline is clearly described and the writing is easy to follow.

The experimental section is thin, even for a late-breaking paper. One dataset, no baselines, no error bars, no code, no held-out configurations. The t-SNE plot and side-by-side images are suggestive but qualitative. The spatial locality check is a nice touch, but it's one figure. The stress-test is right: without an inverse transform, you can't tell whether bursts and inter-arrival times survive the rasterization. Storage workloads are dominated by fine-grained timing; aggregate statistics like read/write ratio and utilization can be preserved while the arrival process is destroyed.

I wouldn't call the 8% error circular—it's a legitimate check that the conditioning mechanism is doing something—but it's not evidence of fidelity. It's the expected outcome if the contrastive embedding works. The real gap is that nothing in the paper demonstrates the generated artifact can be fed into a storage simulator or otherwise used for evaluation.

Who is this for? People working on synthetic workload generation who want to see whether diffusion is a productive direction. It deserves a serious referee because the core idea is worth pinning down; any competent reviewer would ask for the decoding step and a downstream evaluation (e.g., feeding traces into a simulator and comparing performance). If the authors can show that, it could become a real tool. As is, I would not cite it as a reliable generator, but I would definitely read the extended version.

Referee Report

4 major / 4 minor

Summary. The paper proposes DiTTO, a diffusion-based framework for generating synthetic multi-device storage traces. Raw logs are rasterized into multi-channel images (time bins × devices, read/write channels); a contrastively learned encoder CHIP embeds user-specified numeric workload configurations; Gaussian noise augmentation is used to counter sparsity; and outpainting is used to extend traces to arbitrary lengths. Evaluation on Alibaba Block IO traces reports an average error below 8% for read/write ratio and below 2% for device utilization, together with t-SNE and visual comparisons against real traces. The central claim is that DiTTO produces realistic, diverse, and precisely configurable storage traces.

Significance. If fully substantiated, DiTTO would address a real and timely problem: generating controllable, realistic storage traces without expensive instrumentation. The combination of image-based diffusion, contrastive numeric conditioning, sparsity-aware augmentation, and outpainting is original in the storage-trace domain and could provide a useful tool for workload evaluation. However, the current evidence is preliminary. The paper does not specify how the generated continuous image is turned back into a discrete event stream, the headline accuracy numbers are reported without an evaluation protocol, and fidelity/diversity are supported mostly by qualitative figures. The contribution is promising but needs substantial additional validation before it can support the stated claims.

major comments (4)

The paper never describes an inverse transform from the generated continuous image back to a storage trace: there is no thresholding, event-time sampling, or op-type decoding that would yield timestamped read/write requests. Rasterizing time into fixed bins and applying Gaussian smoothing collapses multiple events in one bin and erases exact burst timing. All metrics in Fig. 2 (R/W ratio error, device utilization, t-SNE) are computed on the image representation, not on a discrete event stream usable by a storage simulator. Consequently, the headline '<8% error' quantifies conditioning alignment on images, not event-level fidelity. Please define and validate the image-to-event decoding and report metrics on the recovered event stream (e.g., inter-arrival time distributions, burst lengths, per-device request sequences).
The central quantitative result, an average error of less than 8% for read/write ratios and less than 2% for device utilization, is reported without error bars, number of seeds, train/test split, or held-out configurations. Because the metric is measured against the exact configuration vector used as the conditioning signal, it demonstrates that the model can memorize or regress to its input condition, not that it generalizes to unseen configurations. Please provide a clear evaluation protocol: separate training and test configurations, multiple random seeds, variance across runs, and comparison with a non-conditioned baseline or an existing synthetic trace generator to calibrate how tight 8% is.
The claims of high fidelity and diversity rest on qualitative evidence: a t-SNE plot with overlapping clusters and side-by-side images. t-SNE can show cluster structure but is not a quantitative measure of distributional fidelity or diversity. Please report distributional distances (e.g., Wasserstein distance or MMD on request-rate, spatial-locality, and burstiness features) and a diversity metric (e.g., average pairwise distance among generated traces). Similarly, outpainting is validated only by a single visual example in Fig. 3(c); no quantitative measure of long-horizon coherence or continuity is provided. This is load-bearing because arbitrary-length generation is one of the paper's stated contributions.
The paper claims a "sparsity-aware training approach" and the use of local Gaussian noise augmentation, but gives no details: no augmentation variance, schedule, or training loss modifications. More importantly, the transformation is asserted rather than validated. Storage workloads are characterized by bursty, heavy-tailed arrival patterns; smoothing events into intensity gradients may wash out the fine-grained timing structure that downstream system evaluation depends on. Please provide the concrete augmentation scheme and an analysis showing that the image representation preserves or controllably trades off event-level arrival statistics, especially burstiness and idle-period structure.

minor comments (4)

The manuscript has no related-work section, so the novelty claim relative to [1] and other ML-based trace generators is not clearly positioned. A short discussion of prior statistical models and feature-engineered ML approaches would help readers assess the contribution.
The text contains a long garbled block of hexadecimal-like characters immediately after "implemented DiTTO using Pytorch 2.4". This appears to be a corrupted passage and should be removed or replaced.
The figures are hard to read: Fig. 2(a) lacks units and definitions of "Most accessed" and "Most used ID"; Fig. 3 panels are very small and the read/write channel labels are not clearly legible. Please enlarge and annotate.
The paper does not state how many traces are used for training versus testing, how the 50 generated traces in Fig. 2(a) are sampled, or what subset of the Alibaba dataset is used. These details are necessary to interpret the accuracy numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported config-following errors are direct empirical tests of the learned conditioning mechanism, and the t-SNE and spatial-locality analyses give independent grounding.

full rationale

The paper's derivation chain is not circular. The central quantitative claim is that DiTTO, when conditioned on workload configurations such as read/write ratio and device utilization, generates outputs that match those configurations to within 8% and 2% respectively. This is a direct test of the learned CHIP conditioning mechanism, not a quantity forced by construction: the diffusion model must learn the mapping from configuration embeddings to pixel-level trace patterns, and there is no analytic or fitted relationship that would make the error zero. The t-SNE comparison against the external Alibaba Block IO trace dataset provides independent evidence of realism and diversity, and the spatial-locality cluster experiment evaluates a property that was deliberately not included in the conditioning vector, so it cannot be self-referential. The absence of a described inverse transform from the smoothed continuous image back to discrete timestamped read/write events is a validation gap and a correctness risk, but it is not circularity: the paper never claims that the configuration error is derived from the input configuration. There are no load-bearing self-citations, imported uniqueness theorems, or ansatz-smuggling citations. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that rasterized, Gaussian-smoothed images preserve trace statistics, that the Alibaba dataset represents real workloads, and that diffusion plus contrastive conditioning learns the intended distribution. The model weights and several hyperparameters are free parameters fitted to the dataset, with no sensitivity analysis. There are no invented physical entities; all introduced components are methodological.

free parameters (4)

Gaussian noise augmentation variance
Chosen to smooth sparse event pixels into intensity gradients; no value or sensitivity analysis is given (Section II, stage 1).
Image grid resolution (time bins x devices)
Determines temporal and device granularity of the trace representation; no justification or ablation study (Section II, stage 1).
Diffusion model hyperparameters
Timesteps, U-Net architecture, learning rate, and training epochs are not reported; all results depend on these fitted values (Section III).
CHIP embedding dimension and contrastive learning temperature
No values are given for the contrastive embedding or loss temperature, which control how configuration vectors are aligned with image representations (Section II, stage 2).

assumptions (4)

domain assumption Rasterizing traces into images with Gaussian smoothing preserves the workload statistics needed for system evaluation
The paper presents no validation that information loss from rasterization is acceptable; this is the core representation choice in stage 1 of Section II.
domain assumption Alibaba Block IO trace dataset is representative of real-world multi-device storage workloads
Only this single dataset is used, yet the conclusion generalizes to 'realistic traces' for RAID and CephFS environments (Section III, Section IV).
domain assumption t-SNE visual overlap is sufficient evidence of realism and diversity
Figure 2(b) is the main fidelity evidence beyond visual inspection; no quantitative distance or distribution metric is used (Section III).
standard math Denoising diffusion training objective learns the conditional trace distribution
Standard for diffusion models, but the paper provides no equations or training details, so this is assumed rather than demonstrated (Section II).

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Diffusion-Based Framework for Configurable and Realistic Multi-Storage Trace Generation." pith.science (2026). https://pith.science/paper/P53QMHGA

@misc{pith2026250901919,
  author       = {Pith},
  title        = {Pith review of: A Diffusion-Based Framework for Configurable and Realistic Multi-Storage Trace Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P53QMHGA}},
  note         = {Machine review of arXiv:2509.01919}
}

read the original abstract

We propose DiTTO, a novel diffusion-based framework for generating realistic, precisely configurable, and diverse multi-device storage traces. Leveraging advanced diffusion techniques, DiTTO enables the synthesis of high-fidelity continuous traces that capture temporal dynamics and inter-device dependencies with user-defined configurations. Our experimental results demonstrate that DiTTO can generate traces with high fidelity and diversity while aligning closely with guided configurations with only 8% errors.

Figures

Figures reproduced from arXiv: 2509.01919 by the authors.

**Figure 1.** Overview of DiTTO constrained output control. Additionally, we design a sparsityaware training approach to ensure that long periods of inactivity and bursty access patterns are accurately modeled. Finally, by leveraging outpainting techniques, a generative technique that extends content while preserving contextual coherence, DiTTO produces traces of arbitrary length while maintaining contextual consistency. In this… view at source ↗

**Figure 2.** Generation error and t-SNE 0 255 255 255 Normalized time Device ID READ WRITE 0 0 255 255 255 Normalized time READ WRITE 0 a) Original Trace b) Generated Trace c) Outpainted Trace 255 Normalized time Init Extend Init Extend [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

[1]

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

work page
[2]

Machine learning assisted hpc workload trace generation for leadership scale storage systems

Arnab K Paul, et al. Machine learning assisted hpc workload trace generation for leadership scale storage systems. In Proceedings of the 31st International Symposium on High-Performance Parallel and Distributed Computing , pages 199--212, 2022

work page 2022
[3]

Hierarchical Text-Conditional Image Generation with CLIP Latents , 2022

Aditya Ramesh, et al. Hierarchical Text-Conditional Image Generation with CLIP Latents , 2022

work page 2022
[4]

Learning transferable visual models from natural language supervision, 2021

Alec Radford, et al. Learning transferable visual models from natural language supervision, 2021

work page 2021
[5]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

work page

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

work page

[2] [2]

Machine learning assisted hpc workload trace generation for leadership scale storage systems

Arnab K Paul, et al. Machine learning assisted hpc workload trace generation for leadership scale storage systems. In Proceedings of the 31st International Symposium on High-Performance Parallel and Distributed Computing , pages 199--212, 2022

work page 2022

[3] [3]

Hierarchical Text-Conditional Image Generation with CLIP Latents , 2022

Aditya Ramesh, et al. Hierarchical Text-Conditional Image Generation with CLIP Latents , 2022

work page 2022

[4] [4]

Learning transferable visual models from natural language supervision, 2021

Alec Radford, et al. Learning transferable visual models from natural language supervision, 2021

work page 2021

[5] [5]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

work page