Predicting Amazon Spot Prices with LSTM Networks
Matt Baughman, Christian Haas, Rich Wolski, Ian Foster, Kyle Chard
Abstract
Matt Baughman, Christian Haas, Rich Wolski, Ian Foster, Kyle Chard
Abstract
Amazon spot instances provide preemptable computing capacity at a cost that is often significantly lower than comparable on-demand or reserved instances. Spot instances are charged at the current spot price: a fluctuating market price based on supply and demand for spot instance capacity. However, spot instances are inherently volatile, the spot price changes over time, and instances can be revoked by Amazon with as little as two minutes' warning. Given the potential discount---up to 90% in some cases---there has been significant interest in the scientific cloud computing community to leverage spot instances for workloads that are either fault-tolerant or not time-sensitive. However, cost-effective use of spot instances requires accurate prediction of spot prices in the future. We explore here the use of long/short-term memory (LSTM) recurrent neural networks for spot price prediction. We describe our model and compare it against a baseline ARIMA model using historical spot pricing data. Our results show that our LSTM approach can reduce training error by as much as 95%.
OpenAlex reports 43 citations for this work. Citation counts describe recorded attention and do not establish research quality.
A contribution statement is not available in the OpenAlex record.
Method details are not available in the OpenAlex metadata.
Findings are not separately available in the OpenAlex metadata.
Limitations are not available in the OpenAlex metadata.
Application details are not available in the OpenAlex metadata.
Amazon spot instances provide preemptable computing capacity at a cost that is often significantly lower than comparable on-demand or reserved instances. Spot instances are charged at the current spot price: a fluctuating market price based on supply and demand for spot instance capacity. However, spot instances are inherently volatile, the spot price changes over time, and instances can be revoked by Amazon with as little as two minutes' warning. Given the potential discount---up to 90% in some cases---there has been significant interest in the scientific cloud computing community to leverage spot instances for workloads that are either fault-tolerant or not time-sensitive. However, cost-effective use of spot instances requires accurate prediction of spot prices in the future. We explore here the use of long/short-term memory (LSTM) recurrent neural networks for spot price prediction. We describe our model and compare it against a baseline ARIMA model using historical spot pricing data. Our results show that our LSTM approach can reduce training error by as much as 95%.
Key concepts: Spot contract, Spot market, Leverage (statistics), Autoregressive integrated moving average, Computer science, Blind spot, Hot spot (computer programming), Bright spot