🤖 AI

SWIFT: Spatio-temporal Wavelet Integrated Forecasting Framework for Workload Traces

SWIFT is a pure convolutional framework that enhances cloud workload forecasting accuracy and efficiency by introducing a Learnable Cascaded Wavelet Path for adaptive feature extraction and a Multivariate Interaction Module to model spatial-temporal dependencies, achieving state-of-the-art performance with linear complexity.

Original authors: Zeyuan Ding, Lingfeng Zheng, Dian Ding, Guangtao Xue

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zeyuan Ding, Lingfeng Zheng, Dian Ding, Guangtao Xue

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the traffic flow on a massive highway system (your cloud servers) to know when to open more lanes or close some down. This is the job of workload forecasting.

The problem is that this traffic is chaotic. Sometimes it's a steady stream, but often, huge "bursts" of cars appear out of nowhere at specific times, while other times, the flow follows a predictable daily rhythm.

Current methods for predicting this traffic have two main flaws:

The "Blurry Photo" Problem: Some methods look at the whole day at once to find patterns. It's like taking a long-exposure photo of a highway; you see the light trails, but you can't tell when a specific car sped by. You miss the sudden bursts.
The "Isolated Driver" Problem: Other methods look at each road (server) individually, ignoring that if one road jams, the connected roads usually jam too. They miss the big picture of how the whole system is connected.

Enter SWIFT (Spatio-temporal Wavelet Integrated Forecasting). Think of SWIFT as a super-smart, high-speed traffic analyst that fixes both problems.

How SWIFT Works (The Magic Tricks)

1. The "Peeling Onion" Technique (Learnable Cascaded Wavelet Path)
Imagine you have a complex, messy onion representing the traffic data. Old methods just chopped it up with a dull, fixed knife (rigid math), which often missed the layers or squished the delicate parts.
SWIFT uses a smart, adjustable knife. It peels the onion layer by layer, but here's the trick: the knife learns as it goes.

Layer 1: It peels off the tiny, fast-moving details (the sudden bursts of traffic).
Layer 2: It peels off the medium-sized patterns (hourly rhythms).
Layer 3: It peels off the big, slow trends (daily cycles).
Because the knife is "learnable," it adapts to the specific shape of the onion (the data) rather than forcing a standard cut. This allows it to spot exactly when a traffic burst happens, not just that it happened.

2. The "Group Chat" (Multivariate Interaction Module)
Old methods treated each server like a person talking to themselves in a soundproof room. SWIFT puts them all in a group chat.

Stage 1 (The Group Chat): It lets the servers "talk" to each other first. If Server A sees a spike, it instantly tells Server B, "Hey, we're getting busy!" This stabilizes the noise and helps them understand the system-wide context.
Stage 2 (The Deep Dive): Once they are calm and connected, they analyze their own specific details.
This ensures the model understands that the traffic on one road affects the traffic on the next.

3. The "Race Car" Engine (Pure Convolutional Architecture)
Many powerful prediction models are like heavy, luxury buses. They are accurate but slow and burn a lot of fuel (computing power). They take too long to make a decision, which is bad for real-time traffic management.
SWIFT is built like a Formula 1 race car. It uses a "pure convolutional" design, which is a type of math operation that modern computer chips (GPUs) love. It's incredibly fast, linear in speed (meaning if you double the data, it only takes double the time, not four times), and doesn't get bogged down.

The Results: Why It Matters

The paper tested SWIFT on real-world data from massive cloud providers (like Alibaba and Google) and found:

It's More Accurate: It reduced prediction errors by up to 31% compared to the best existing methods. It's much better at spotting those sudden traffic bursts before they cause a jam.
It's Much Faster: It cut the time it takes to make a prediction by nearly 80%.
It's Efficient: It uses less computer memory and power while still being the most accurate.

In a Nutshell

SWIFT is a new way to predict cloud server traffic. Instead of using blurry, slow, or isolated methods, it uses a smart, layer-by-layer peeling technique to spot sudden changes and a group chat system to understand how different servers affect each other. The result is a prediction tool that is both sharper (more accurate) and faster (more efficient) than anything currently available.

Technical Summary: SWIFT - Spatio-temporal Wavelet Integrated Forecasting Framework for Workload Traces

1. Problem Statement

Accurate cloud workload forecasting is essential for proactive resource management (e.g., scaling and scheduling) but faces significant challenges due to the nature of real-world cloud traces. The authors identify two primary hurdles:

Complex Workload Patterns: Cloud workloads are highly volatile, characterized by sudden, localized bursts interlaced with multi-scale periodicities (daily, hourly). Furthermore, microservice dependencies create strong spatial coupling, causing synchronized fluctuations across different workload traces.
- Limitation of Existing Methods: Traditional Fourier-based methods (e.g., FFT) utilize global basis functions, which aggregate energy over the entire time domain. This results in a loss of temporal locality, making it impossible to distinguish when specific bursts occur (leading to phase lags and untimely reactions). Conversely, existing wavelet-based methods often rely on rigid, fixed decomposition schemes that struggle with intricate multi-scale patterns and fail to model the spatial dependencies between coupled variables.
Extreme Efficiency Requirements: Large-scale platforms (e.g., ByteDance) require processing over 100,000 concurrent prediction tasks per hour. High-performance models like Transformers suffer from quadratic $O(L^2)$ complexity, causing inference latency and memory exhaustion. Meanwhile, standard Convolutional Neural Networks (CNNs) often operate solely in the time domain, neglecting spatial correlations, while traditional wavelet implementations are sequential and unparallelizable on GPUs.

2. Methodology

The authors propose SWIFT (Spatio-temporal Wavelet Integrated Forecasting Framework), a pure convolutional framework designed to synergize multi-resolution wavelet analysis with efficient convolutional architectures. The framework consists of two core modules:

A. Learnable Cascaded Wavelet Path (LC-Wavelet)

This module addresses the limitations of fixed wavelet bases and sequential processing.

Convolutional Re-engineering: Instead of recursive CPU-based algorithms, SWIFT reformulates the Discrete Wavelet Transform (DWT) and Inverse DWT (IWT) as parallelizable depth-wise convolution operations. This allows native execution on GPU Tensor Cores, eliminating CPU-GPU context switching.
Cascaded Feature Peeling: The architecture employs a deep, cascaded structure where the low-frequency component from one level is recursively fed as input to the next. This creates an exponential expansion of the receptive field, enabling the model to progressively isolate high-frequency stochastic bursts in shallow layers while preserving long-range trends in deeper layers.
Learnable Coefficient Rectification: To overcome the rigidity of fixed bases, the model introduces a two-step learnable mechanism at each decomposition level:
1. Morphological Feature Correction: A learnable convolution layer transforms static coefficients into adaptable representations, smoothing artifacts and aligning features with actual workload shapes.
2. Adaptive Spectral Filtering: A channel-wise scaling vector dynamically re-weights frequency components, suppressing stochastic noise and amplifying informative periodic signals.

B. Multivariate Interaction Module (MIM)

This module captures spatial correlations between variables without incurring the quadratic complexity of attention mechanisms.

Two-Stage Design: MIM operates purely via grouped convolutions.
1. Inter-variable Spatial Interaction: Uses grouped $1 \times 1$ convolutions (groups = feature dimension) to mix information across the variable dimension. This leverages system-wide context to stabilize noisy workload states before feature extraction.
2. Intra-variable Feature Interaction: Uses grouped convolutions (groups = variable dimension) to enable dense, non-linear information exchange within each variable's feature embedding.
Efficiency: By decoupling spatial mixing and feature mixing, MIM captures the comprehensive state of the cloud system with linear $O(L)$ complexity.

3. Key Contributions

Learnable Cascaded Wavelet Path: A novel architecture that transforms the rigid, fixed-basis wavelet transform into a deep, data-driven, and fully parallelizable convolutional operator, enabling precise multi-scale feature decoupling.
Multivariate Interaction Module (MIM): A spatial modeling component that captures inter-variable dependencies and intra-variable feature interactions using grouped convolutions, bridging the gap between isolated decomposition and system-wide coupling.
Hardware-Friendly Pure Convolutional Architecture: A design that replaces recursive CPU calls with GPU-native depth-wise convolutions, achieving linear $O(L)$ complexity while eliminating data transfer bottlenecks.

4. Experimental Results

The authors evaluated SWIFT on four real-world cloud workload datasets (Ali2021, Ali2022, Fisher, Google) and the ETT benchmarks.

Accuracy: SWIFT achieved State-of-the-Art (SOTA) performance in 7 out of 8 long-term forecasting cases. It reduced prediction errors by up to 31.04% in short-term forecasting (SMAPE) and up to 20.05% in generalization scenarios compared to baselines.
Efficiency: SWIFT demonstrated superior inference speed, reducing latency by 79.74% compared to CNN-based baselines and 66.82% compared to Transformer-based baselines. It maintained linear complexity while achieving high throughput.
Generalization: In zero-shot cross-dataset experiments (training on Ali2022, testing on others without fine-tuning), SWIFT maintained SOTA results, demonstrating its ability to capture universal multi-scale periodicities rather than overfitting to specific domain statistics.
Ablation Studies: Removing the LC-Wavelet path caused the most severe performance degradation, confirming it as the core backbone. Removing the MIM module also significantly increased errors, validating the necessity of modeling spatial dependencies.

5. Significance and Claims

The paper claims that SWIFT successfully bridges the gap between precise pattern capture and deployment efficiency. By replacing rigid Fourier/Wavelet schemes with learnable, parallelizable convolutions, the framework addresses the "indistinguishable spectra" problem of FFT and the "spatial neglect" of standard CNNs.

The authors assert that SWIFT enables proactive auto-scaling by providing high-accuracy, low-latency forecasts capable of reacting to sudden workload bursts before they occur. Beyond cloud computing, the authors suggest these capabilities could extend to edge computing, energy grids, and logistics, though they caution that improved accuracy does not guarantee stability and recommend fail-safe mechanisms for automated scaling in unseen anomaly scenarios.