Measuring the Variation of Function Invocation for Cold Start Optimization in Serverless Environments

The cold start issues are common in the Serverless environment, where the predictability of function invocations is regarded as a critical factor for addressing such issues. Many previous studies design specific optimization strategies via reinforcement learning (e.g, Q-learning algorithm), but they do not fully consider the differences in the predictability of function invocations. In the Serverless environment, some functions exhibit higher predictability with fixed patterns that can be well predicted, therefore the optimal scheduling strategy can be determined. As a contrast, some other functions have lower predictability as the invocations are completely random, and these function invocations significantly restrict the effectiveness of reinforcement learning for finding a better scheduling strategy toward cold start issues.To address this problem, in this paper, we propose a novel strategy for cold start optimization in Serverless environment. Specifically, we incorporate the coefficient of variation (CV) into the Q-learning algorithm so as to measure the variation of function invocations, based on which the predictability of functions can be more accurately estimated, thus the optimization strategy for cold start issues can be determined accordingly. Experiments based on the Microsoft Azure dataset verify the effectiveness of the proposed optimization strategy, where the effect of the CV metric is highlighted.