Tuesday 08 April 2025
As our daily lives become increasingly dependent on technology, predicting how long it will take for a computer program or app to complete its task is becoming more and more important. This problem is particularly challenging in edge computing, where devices are limited by their processing power and memory. A new approach has been developed that takes into account the interference between different programs running on these devices, allowing for more accurate predictions.
The researchers behind this project collected data from a variety of devices and WebAssembly runtimes to train a machine learning model called Pitot. This model uses a combination of workload and platform features, such as opcode counts and CPU architecture, to predict runtime. But what makes Pitot different is its ability to account for interference between workloads.
In traditional machine learning models, interference is often ignored or handled separately. However, in edge computing, it’s crucial to consider how multiple programs will impact each other’s performance. Pitot’s authors achieved this by introducing an interference matrix that captures the susceptibility of each platform to interference from other workloads.
The results are impressive. When predicting runtime without interference, Pitot outperforms traditional machine learning models. But when interference is present, it’s able to accurately capture its impact and make more accurate predictions. In some cases, this means reducing the error by as much as 50%.
To better understand how Pitot works, the researchers visualized the learned embeddings of workload and platform features using a technique called t-SNE. These visualizations show that similar workloads cluster together, while platforms with different WebAssembly runtimes or CPU architectures form distinct groups.
The interference matrix also provides valuable insights into the characteristics of each platform. By analyzing its spectral norm, researchers can identify which devices are most susceptible to interference and how much it will impact their performance.
Pitot’s implications extend beyond edge computing. Its ability to account for interference has far-reaching potential in fields such as cloud computing, where multiple workloads often compete for resources. By improving prediction accuracy, Pitot could help optimize resource allocation and reduce latency.
As technology continues to advance, the need for accurate runtime predictions will only grow more pressing. With Pitot, researchers have taken a crucial step towards solving this challenge. Its unique approach to interference modeling holds great promise for improving the performance and efficiency of edge computing devices.
Cite this article: “Predicting Edge Computing Runtime in Presence of Interference: A Conformal Matrix Completion Approach”, The Science Archive, 2025.
Edge Computing, Machine Learning, Runtime Prediction, Webassembly, Interference Modeling, Workload, Platform Features, Opcode Counts, Cpu Architecture, T-Sne, Spectral Norm
