acceptodds
Under review as a conference paper at ICLR 2027

M3-DD: Detecting and Localizing Multiple Concurrent Performance Degradations in Cloud Systems

Abstract

Performance degradation is one of the most pressing threats to the reliability of modern cloud systems, yet its root causes are latent, entangled, and rarely confined to a single metric. Many available diagnostic benchmarks reduce degradation to binary or single-fault labels rather than multiple time-bounded events that may overlap. In this paper, we present the Multi-modal Concurrent Degradation Dataset (MCDD), a VM-level benchmark assembled from production observations, with 22 metric channels, a vocabulary of 122 low-level anomaly-event types, and interval annotations for 25 degradation categories. These annotations preserve multiple overlapping events. We further propose M-DD (Multi-Modal Multi-instance Degradation Detector), which uses object queries to jointly classify and localize overlapping degradation intervals. Its interval-aware attention fuses metric time series with low-level anomaly events, allowing the decoder to represent their temporal relationships before predicting event sets. We evaluate native concurrent event sets using Coverage–Localization Balance (CLB), a composite of event-level alert coverage and localization precision, alongside confidence-ranked average mAP. On MCDD, M-DD attains the highest CLB (0.4180) and average mAP (0.1967) among the evaluated native-view methods. These results highlight M-DD’s ability to detect and localize concurrent degradations.

Primary area:
learning on time series and dynamical systems
Keywords:
cloud performance degradation, multi-modal time series, concurrent faults, temporal action localization

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_Sd1EOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces MCDD, a production-derived VM-level dataset for concurrent degradation detection, and M³-DD, a multi-modal multi-instance detector. MCDD contains 22 metric channels, 122 low-level anomaly-event types, and 25 degradation categories annotated as overlapping, time-bounded intervals. M³-DD combines an anomaly-event encoder, a temporal-signal encoder, interval-aware attention Transformer (IAAT), cross-modal fusion, and an object-query decoder with Hungarian set matching.

Critical strengths
  1. Existing cloud/microservice benchmarks mostly use binary or point labels, single root causes, or window-level annotations. MCDD preserves overlapping, typed, time-bounded degradation events per VM, which is closer to practice diagnosis.
  2. M³-DD combines multiple-instance learning with temporal-boundary regression to identify the categories and recorded intervals of overlapping faults.
Critical weaknesses
  1. The constructed dataset lacks a considerable amount of details. For instance, there is no detailed description of the 25 degradation categories, and the full paper also lacks visualizations related to the dataset.
  2. This method utilizes low-level metric-derived anomaly events as auxiliary inputs. In real-world scenarios, where these data should be obtained?
  3. The references are completely outdated and seriously lack research conducted in 2025-2026.
Questions for authors
  1. Why low-level metric-derived anomaly events can be utilized in an anomaly task scenario? Although these annotations are present in the benchmark proposed in this paper, in real-world scenarios, these data should all be missing.
  2. Are the auxiliary inputs also required during testing?
  3. Given that no additional materials were provided, could a complete and detailed explanation to the dataset be offered?
Reviewer_49pwOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper focuses on detecting and localizing multiple concurrent performance degradations in cloud systems. It first introduces a new multimodal dataset containing 22 monitoring metrics and 122 low-level anomaly event types as inputs, with 25 degradation categories as prediction targets. The paper then proposes a new method for this task, which combines multimodal features through Interval-Aware Attention (IAAT) and uses a DETR-style query decoder to predict degradation types and time intervals. Experiments on the proposed benchmark and four public benchmarks show improvements over several existing methods.

Critical strengths
  1. The problem of detecting multiple overlapping performance degradations is practically relevant.

  2. The experiments cover several baselines and include ablation studies and cross-dataset evaluations.

Critical weaknesses
  1. The paper is difficult to follow, and the methodological design is not well motivated. Many important concepts and details are ambiguous or insufficiently explained. The main paper is not sufficiently self-contained, as too many essential details, such as the descriptions of baseline methods and dataset construction, are deferred to the appendix. Moreover, although adopting existing techniques is reasonable for application-oriented research, the paper does not adequately motivate or justify its design choices. Instead, much of the methodological discussion focuses on describing existing techniques, making it difficult to assess the novelty and contributions of this work.

  2. The experiments do not sufficiently validate the key methodological contributions. While the paper proposes multimodal fusion and IAAT to address concurrent degradation detection, the evaluation mainly demonstrates improvements in overall performance without establishing how these designs specifically address the challenges of overlapping events. More targeted experiments are needed to justify the proposed design choices and demonstrate their effectiveness for the intended task.

  3. Additional issues with formatting and presentation:

    • Several citations in the Related Work section are incorrectly formatted.

    • The anonymous code repository linked in the reproducibility statement was empty at the time of review.

    • Table 1 would be clearer if the explanatory sentence below the table were moved to the caption or main text.

Questions for authors
  1. What are the 25 performance degradation categories, and how do they differ from the 122 low-level anomaly event types? Could the authors provide concrete examples illustrating the inputs, ground-truth annotations, and expected predictions?

  2. The term "concurrent" is potentially misleading, as the task appears to focus on detecting temporally overlapping or interleaved degradation events rather than concurrency in the conventional systems sense. Could the authors clarify what "concurrent" means in this context and how it differs from simply detecting overlapping events?

Reviewer_RDtZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds MCDD, a VM-level dataset with 22 metric channels, 122 types of low-level anomaly events, and interval labels for 25 degradation types, where several labels can overlap in time. It then proposes M3-DD, which encodes the metrics and the events with an interval-aware version of TAAT (IAAT), joins the two by concatenation, and predicts a set of typed intervals with a DETR-style query decoder. M3-DD gets the best CLB (0.418) and Avg mAP (0.197) on MCDD.

Critical strengths

Labeling degradations as typed, overlapping intervals is closer to what operators see than binary or window-level labels. The paper also reports coverage and localization separately, which makes the numbers easier to read.

Critical weaknesses
  1. The data source is not described well enough. The paper only says the data come from "production VM observations" and "aggregated production diagnostic records". It does not say which provider or platform, how the VMs and time range were chosen, or how many labels come from rules, customer feedback, and expert fixes. If many labels come from rule-based diagnosis over the same metrics that produce the 122 input events, the model may partly learn to copy those rules. This also matters for the strong event-only result in Table 3. The data are not released for review, so these points cannot be checked.

  2. The meaning of "concurrent fault" is narrow. In MCDD, a concurrent case is several single-cause faults whose intervals overlap. In practice, a more common and harder case is one degradation caused by several factors acting together (for example, network retransmission and CPU contention that jointly raise latency). The paper does not separate these two cases, and Section 2.2 is not clear whether "concurrent" means multi-dimensional (several fault types) or multi-variable (several metrics behind one fault). The claims in the introduction should be limited to the first case, or the second case should be studied.

  3. There is no part of the model that deals with the relations among faults. Each query predicts one event on its own, and the set matching loss treats each ground-truth event separately. Nothing in the model or the loss uses how one fault affects another, or how a group of metrics points to one fault. Since the paper is about concurrent degradations, this is the part I expected to see.

  4. The dataset is small for this task. There are 2,744 labeled intervals over 25 types (about 110 per type on average, with 1,711 in training), and the test set has 389 VMs. With so few samples per type, it is hard to know whether training is stable. The paper reports one run per setting, with no seeds, variance, or per-class results.

Questions for authors

Which platform do the data come from, and what share of the 2,744 labels comes from rule-based diagnosis? Do these rules use the same detectors that produce the 122 input event types?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.