摘要：This issue of tweets will introduce 3.1 Feature extraction and 3.2 Temporal granularity attention module of "Elasticity unleashed:

分享兴趣，传播快乐，

增长见闻，留下美好。

亲爱的您，这里是LearingYard学苑！

今天小编为您带来“越览（181）——精读期刊论文

《Elasticity unleashed: Fine-grained cloud scaling

through distributed three-way decision

fusion with multi-head attention》的

3.1特征提取和3.2时间粒度注意力模块”。

欢迎您的访问！

Share interest, spread happiness,

increase knowledge, and leave beautiful.

Dear, this is the LearingYard Academy!

Today, the editor brings the

＂Yue Lan (181)：Intensive reading of the journal article

＇Elasticity unleashed: Fine-grained cloud scaling

fusion with multi-head attention’

3.1 Feature extraction and

3.2 Temporal granularity attention module.

Welcome to visit!

一、内容摘要（Summary of Content）

本期推文将从思维导图、精读内容、知识补充三个方面介绍精读期刊论文《Elasticity unleashed: Fine-grained cloud scaling through distributed three-way decision fusion with multi-head attention》的3.1特征提取和3.2时间粒度注意力模块。

This issue of tweets will introduce 3.1 Feature extraction and 3.2 Temporal granularity attention module of "Elasticity unleashed: Fine-grained cloud scaling through distributed three-way decision fusion with multi-head attention" from three aspects: mind map, intensive reading content, and knowledge supplement.

二、思维导图（Mind map）

三、精读内容（Intensive reading content）

（一）特征提取（Feature extraction）

本节主要介绍了研究中使用Google Cluster Trace数据集的方法和分析思路。

This section mainly introduces the methods and analysis ideas of using the Google Cluster Trace dataset in this study.

1. 数据集介绍（Dataset introduction）

本研究采用了 Google Cluster Trace 数据集，该数据集涵盖了大规模云计算集群中作业与任务的详细运行情况，从作业到达时间到 CPU 与内存使用率等多种属性均有完整记录，能够真实反映云环境下任务的动态行为。这些信息对于探索工作负载规律、识别性能瓶颈和优化。资源利用具有重要价值，因此成为后续分析与建模的基础。

This research used the Google Cluster Trace dataset, which provides detailed information about the execution of jobs and tasks in large-scale cloud computing clusters. This dataset comprehensively records various attributes, from job arrival times to CPU and memory usage, truly reflecting the dynamic behavior of tasks in cloud environments. This information is crucial for exploring workload patterns, identifying performance bottlenecks, and optimizing resource utilization, forming the foundation for subsequent analysis and modeling.

2. 核心分析指标（Core analysis indicators）

尽管 Google Cluster Trace 提供了丰富的特征，但本研究的重点聚焦在CPU 利用率和内存使用率两个关键指标。CPU利用率衡量处理器在单位时间内的工作负载，高值意味着系统繁忙，低值则代表资源闲置；内存使用率反映系统内存的占用情况，持续监控这一指标有助于避免内存瓶颈并提升整体资源分配效率。为了便于量化分析，文中对两个指标进行了数学定义，其中Ntasks表示任务数量，ti表示任务的开始时间。

While Google Cluster Trace provides a rich set of features, this study focuses on two key metrics: CPU utilization and memory usage. CPU utilization measures the processor workload per unit time; high values indicate a busy system, while low values indicate idle resources. Memory usage reflects system memory usage. Continuously monitoring this metric helps avoid memory bottlenecks and improve overall resource allocation efficiency. To facilitate quantitative analysis, this paper provides mathematical definitions for these two metrics: Ntasks represents the number of tasks, and ti represents the start time of a task.

3. 时间聚合法（Time aggregation method）

为了更好地捕捉任务运行的动态特征，研究采用了时间聚合方法。具体而言，将原始数据按照不同时间间隔（如 5 分钟、10 分钟、天或周）进行划分，并在每个时间段内计算平均CPU和内存使用率。聚合后的数据被进一步组织成矩阵形式，其中行对应不同的时间区间，列则表示 CPU 与内存两个核心特征。该方法不仅能够揭示不同时间尺度下的工作负载模式，还为后续的模式挖掘和预测提供了结构化数据支持。

To better capture the dynamic characteristics of task execution, the study employed a temporal aggregation approach. Specifically, the raw data was partitioned into time intervals (e.g., 5-minute, 10-minute, daily, or weekly), and the average CPU and memory usage was calculated for each time interval. The aggregated data was further organized into a matrix, with rows corresponding to different time intervals and columns representing the two core characteristics of CPU and memory. This approach not only reveals workload patterns across different time scales but also provides structured data support for subsequent pattern mining and prediction.

4. 分布式三支决策（Distributed three-way decision making）

在有限对象集合U上定义决策集合D={d1,d2,…,dn}，每个 di代表一个粒度（如不同时间间隔）。

Define a decision set D={d1,d2,…,dn} on a finite object set U, where each di represents a granularity (such as a different time interval).

在每个粒度下，三支决策Δi包含三个互斥子集：

At each granularity, the three-way decision Δi contains three mutually exclusive subsets:

分布式三支决策Δ将多个粒度的决策结合起来，实现综合且自适应的决策过程。

Distributed three-way decision Δ combines decisions at multiple granularities to achieve a comprehensive and adaptive decision-making process.

融合方式既可以依赖预设规则，也可以采用复杂算法（如多头注意力机制），从不同粒度智能地整合决策，从而得到最优且具备上下文感知的结果。

The fusion method can rely on preset rules or use complex algorithms (such as multi-head attention mechanism) to intelligently integrate decisions at different granularities to obtain optimal and context-aware results.

（二）时间粒度注意力模块（Temporal granularity attention module）

本节主要介绍了时间粒度注意力模块的机制和作用。

This section mainly introduces the mechanism and function of the temporal granularity attention module.

1. 作用与目标（Function and goal）

时间粒度注意力模块的核心任务是从不同时间尺度（如小时、天、周）中识别和捕捉时序模式。不同粒度的时间序列特征输入后，会被分别处理，以提取出在该时间层次上的关键信息。

The core task of the temporal attention module is to identify and capture temporal patterns at different time scales (e.g., hours, days, and weeks). Time series features at different granularities are input and processed separately to extract key information at that temporal level.

2. 向量变换（Vector transformation）

对每个时间粒度的输入特征向量xt，通过线性变换分别得到查询向量qt、键向量kt、值向量vt。这些向量为后续注意力计算做准备。

For each time granularity input feature vector xt, a linear transformation is performed to obtain the query vector qt, key vector kt, and value vector vt. These vectors prepare for subsequent attention calculations.

3. 注意力权重计算（Attention weight calculation）

通过计算查询向量和键向量的点积，并对结果应用 softmax 函数，得到注意力权重at。该权重表示输入序列中不同元素在该时间粒度下的重要性。

By calculating the dot product of the query vector and the key vector and applying the softmax function to the result, we get the attention weight at. This weight represents the importance of different elements in the input sequence at that time granularity.

4. 输出生成（Output generation）

注意力权重与值向量进行逐元素相乘，得到该粒度下的注意力输出向量ot。这一输出能够突出关键时间点或模式。

The attention weight is multiplied element-by-element by the value vector to obtain the attention output vector ot at that granularity. This output can highlight key time points or patterns.

四、知识补充（Knowledge supplement）

时间粒度注意力机制的提出，源于深度学习中 多头注意力机制的发展。传统的时间序列建模方法，如 ARIMA、LSTM，往往依赖单一时间尺度的信息，容易忽略跨尺度的时序关联。而在实际的云计算、金融预测或能源调度等场景中，任务负载往往同时受到短期波动（小时级别）、中期趋势（天级别）以及长期周期性（周或月级别）的共同影响。单一粒度的建模很难全面刻画这种复杂性。

The proposal of the temporal-granularity attention mechanism stems from the development of the multi-head attention mechanism in deep learning. Traditional time series modeling methods, such as ARIMA and LSTM, often rely on information at a single time scale and tend to overlook cross-scale temporal correlations. However, in real-world scenarios such as cloud computing, financial forecasting, and energy scheduling, workloads are often influenced by short-term fluctuations (hourly), medium-term trends (daily), and long-term cyclical effects (weekly or monthly). Single-granularity modeling struggles to fully capture this complexity.

通过引入时间粒度注意力模块，可以在不同尺度上同时学习特征表示。具体而言，查询、键、值的线性变换使得模型能够在每个粒度下捕捉序列内部的依赖关系；而 softmax 权重的引入，则确保模型能够突出在当前粒度下更为关键的时刻或模式。最终得到的注意力输出向量不仅包含了各时间段的核心特征，还能在多粒度之间形成互补。

By introducing a temporal attention module, feature representations can be learned simultaneously at different scales. Specifically, the linear transformation of queries, keys, and values enables the model to capture dependencies within the sequence at each granularity. The introduction of softmax weights ensures that the model emphasizes the most critical moments or patterns at the current granularity. The resulting attention output vector not only captures the core features of each time period but also complements the multiple granularities.

值得注意的是，这种机制与 多尺度建模思路密切相关。多尺度方法强调在不同层次或分辨率下同时建模，常见于信号处理、小波分析和图像识别中。时间粒度注意力可以看作是这一思想在时间序列和深度学习框架中的延伸，它结合了注意力机制的灵活性和多尺度分析的全面性。其优势在于，能够动态调整模型对不同时段的关注程度，从而在复杂环境下实现更精细化的预测与决策。

It's worth noting that this mechanism is closely related to the concept of multi-scale modeling. Multi-scale methods emphasize simultaneous modeling at different levels or resolutions and are commonly used in signal processing, wavelet analysis, and image recognition. Temporal attention can be seen as an extension of this concept within time series and deep learning frameworks, combining the flexibility of attention mechanisms with the comprehensiveness of multi-scale analysis. Its advantage lies in its ability to dynamically adjust the model's focus on different time periods, thereby achieving more refined predictions and decision-making in complex environments.

今天的分享就到这里了。

如果您对文章有独特的想法，

欢迎给我们留言，让我们相约明天。

祝您今天过得开心快乐！

That's all for today's sharing.

If you have a unique idea about the article,

please leave us a message,

and let us meet tomorrow.

I wish you a nice day!

文案|yyz

排版|yyz

审核|hzy

翻译：谷歌翻译

参考资料：百度百科、Chat GPT

参考文献：Jiang C, Duan Y. Elasticity unleashed: Fine-grained cloud scaling through distributed three-way decision fusion with multi-head attention [J]. Information Sciences, 2024, 660(1): 1-15.

本文由LearningYard学苑整理发出，如有侵权请在后台留言！

来源：LearningYard学苑