Abstract: High-performance computing(HPC) clusters play a crucial role in large-scale computing tasks,however,as the scale of clusters continues to expand,the management of cluster energy efficiency has become increasingly important.Prediction technology can provide decision support for energy efficiency management through accurate estimation of cluster resource usage,thereby achieving dynamic optimization of resources and effective control of energy consumption.This paper proposes an HPC node prediction method called iDSRformer,which introduces a multi-head sparse self-attention mechanism to improve computational efficiency and capture dependencies between multiple features,uses RMSNorm normalization to enhance stability,and adopts depthwise separable convolution to implement the feed-forward network,expanding the model's receptive field to achieve more efficient feature extraction and handle complex relationships between multiple variables.Experiments are conducted on Microsoft's Philly cluster and Alibaba's cluster-trace-gpu-v2020 datasets,respectively.The results show that compared with the currently proposed prediction models iTransformer,Transformer,patchTST,dlinear,timeMixerh,and crossformer,iDSRformer achieves average MSE reduction of 10.2%,20.2%,12.2%,19.2%,10.1%,and 13.3% and average MAE reduction of 9.9%,26.3%,10.8%,16.9%,7.4%,and 15% on the Philly task data,while on Alibaba's cluster-trace-gpu-v2020 task data,it achieves average MSE reduction of 10.2%,18.3%,16%,28.2%,15.4%,and 14.4% and average MAE reduction of 8.8%,14.4%,11.5%,19.5%,9%,and 9%,demonstrating better prediction accuracy.

Key words: Cluster energy saving, Attention mechanism, Time series forecasting, iTransformer, Multi-step forecasting