Core Concepts
PePNet, a novel workload prediction network, improves overall and heavy workload prediction accuracy by leveraging a Periodicity-Perceived Mechanism and an Achilles' Heel Loss Function.
Abstract
The paper proposes PePNet, a workload prediction network that addresses the challenges of highly variable workloads with rare occurrences of heavy workload.
Key highlights:
Periodicity-Perceived Mechanism: PePNet automatically detects the existence and length of periodicity in the workload time series, and adaptively fuses the periodic information to handle both periodic and aperiodic workloads.
Achilles' Heel Loss Function: PePNet uses a specialized loss function that iteratively optimizes the most under-fitting part of the prediction sequence, effectively improving the accuracy of heavy workload prediction while maintaining high overall prediction accuracy.
Extensive experiments on Alibaba, SMD, and Dinda's datasets show that PePNet outperforms state-of-the-art methods, improving heavy workload prediction accuracy by 21.0% and overall prediction accuracy by 11.8% on average.
PePNet introduces only slight time overhead compared to efficient baseline models, while being insensitive to hyperparameter settings.
Stats
"The heavy workload is defined as the workload greater than the average workload plus one standard deviation for each machine."
"The heavy workload proportion is 15.73% for CPU usage and 12.89% for memory usage of Alibaba's dataset, and 11.15% for workload of Dinda's dataset."
Quotes
"Either the overall inaccuracy of statistic methods or the heavy-workload inaccuracy of neural-network-based models can cause service level agreement violations."
"The Achilles' Heel Loss Function actually forces the loss to descend along the gradient directions of some of the largest prediction errors."