Contemp: Instance Caching Based on Container Temperature in Serverless Environment

In serverless computing, cloud providers dynamically manage the underlying resource allocation, allowing developers to concentrate on business logic development. Although containers ensure runtime consistency, creation and destruction of instances significantly degrade performance by increasing cold start latency. A widely adopted solution is instance caching, which facilitates rapid container reuse upon request arrival. However, existing caching mechanisms exhibit limitations under highconcurrency burst workloads: the lack of a dynamic evaluation mechanism for real-time container activity levels may lead to premature eviction of high-activity containers or excessive resource occupation by low-activity containers, thereby reducing resource efficiency and service quality. This paper presents Contemp, a container temperature-based cache eviction strategy. Contemp constructs a Historical-Enhanced Adaptive Temperature Model (HEATM), which integrates temporal patterns, cold start latency, resource efficiency, and historical invocation frequency. In addition, it adopts a decay mechanism based on Newton's Law of Cooling to prevent excessive caching of containers. Based on the temperature awareness of container, Contemp dynamically adjusts cache priorities to guide cache eviction decisions, and achieves a better trade-off between resource costs and cold start latency. Experimental evaluations on Azure Functions trace demonstrate Contemp's superior performance: improving an average of 17.3% higher cache hit rate of each node, 30.1% reduction in cold start latency, and 25.4% cost effectiveness improvement compared to existing approaches.