Computing Cluster Load Prediction from Time-Series Data

Abstract

we studied short-term computing cluster load prediction using the Alibaba GPU Trace 2020 production trace. We considered hourly processor, graphics accelerator, and random-access memory (RAM) loads, with a one-hour prediction horizon and a 168-hour test window. We aggregated task events into hourly time series by averaging the concurrent load over time. We compared five prediction models: a naive last-value model, an integrated moving-average model, a long short-term memory (LSTM) recurrent neural network, a Bayesian dynamic linear model, and a multivariate vector autoregression model with lag selection up to 168 hours. We evaluated prediction accuracy, 99% prediction-interval coverage, accuracy in predicting the direction of load changes, and the potential savings in reserved resources. The Bayesian dynamic linear model achieved the lowest errors, with mean absolute errors 6–9% lower than those of the naive model. The LSTM model was competitive for RAM, whereas the multivariate model performed well for processor and accelerator loads. Replacing reservation based on the historical maximum with reservation based on the upper bound of the prediction interval reduced the reserve by 22–30% while maintaining 96–99% coverage.