Representation Learning for Resource Usage Prediction

Florian Schmidt, Mathias Niepert, Felipe HuiciarXiv 2018 · arXiv

Introduction

Modern operating systems and cloud orchestrators, such as Kubernetes, typically rely on static heuristics for resource allocation. These "one-size-fits-all" rules often fail to handle the dynamic nature of "spiky" workloads or complex, multi-process applications. A fundamental limitation is that high-level telemetry—such as aggregate CPU or memory usage—lacks the semantic context of application behavior. Without knowing whether a process is performing intensive I/O or heavy computation, schedulers cannot accurately forecast future resource demands.

Researchers at NEC Laboratories Europe propose a novel approach that treats sequences of system calls (syscalls) as a "language" of application behavior. By applying representation learning to these syscall streams and combining them with traditional metrics, they enable more accurate, context-aware resource usage prediction.

Methodology

The proposed framework utilizes a multi-modal input pipeline to capture both the physical state and the behavioral intent of a running application.

Architecture of the resource usage prediction pipeline, combining telemetry data with system call embeddings.

The approach consists of three primary stages:

  1. Telemetry Collection: Standard resource metrics, including CPU utilization, Resident Set Size (RSS) memory, and Disk I/O, are sampled over discrete time intervals.
  2. System Call Embeddings: To capture behavioral state, the system monitors the stream of syscalls. Treating these discrete events like words in a sentence, a word2vec skip-gram model is applied to generate fixed-size vector embeddings. This allows the model to learn that specific sequences of calls (e.g., open, read, mmap) represent distinct high-level operations.
  3. LSTM Modeling: The usage statistics and behavioral embeddings are concatenated and fed into a Long Short-Term Memory (LSTM) neural network. The LSTM maintains an internal hidden state to model temporal dependencies, ultimately predicting resource usage at a future time step $t + i$.

Experimental Setup

The model was evaluated using a custom-built scientific computing toolchain. The dataset comprised various bash and Python scripts designed to simulate real-world data center workloads by interleaving phases of heavy network/disk I/O with intensive CPU computation. This variety ensures the model is tested against shifting application behaviors.

Results

Preliminary results focused on predicting scaled CPU usage revealed critical insights into the role of temporal context.

Prediction error (RMSE) as a function of the prediction horizon for different history lengths.

As shown in the evaluation, the prediction error (RMSE) naturally increases as the model attempts to forecast further into the future (up to 15 seconds). However, the amount of historical context provided to the model is a decisive factor. With only 1 second of history, the RMSE climbs sharply as the prediction horizon expands. In contrast, providing 10 seconds of history results in a significantly lower and more stable error rate. This suggests that capturing long-term behavioral patterns through syscall embeddings is essential for robust forecasting.

Discussion and Future Work

While promising, this work identifies several avenues for further research. The computational overhead of real-time syscall tracing and ML inference must be rigorously analyzed for production environments, though tools like perf may mitigate these costs. Additionally, future studies should investigate how well syscall embeddings generalize across diverse software categories, such as web servers versus databases. The ultimate goal is the end-to-end integration of these predictions into OS schedulers or cloud orchestrators to optimize performance and energy efficiency.

Summary updated

BibTeX
@misc{schmidt2018representation,
  author = {Florian Schmidt and Mathias Niepert and Felipe Huici},
  title = {{Representation Learning for Resource Usage Prediction}},
  year = {2018},
  eprint = {1802.00673},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1802.00673},
}
Figure